A comparison of alternative methodologies for estimation of HIV incidence

This dissertation examines two models for estimation of the incidence of HIV infection. One depends on data more commonly found in wealthy countries---STARHS or testing for recent infection---and the other uses repeated prevalence surveys that are the most common source in resource-limited areas. The goals of the project are to improve both models, but also to compare their performance, an exercise that has not been done and will contribute to the ongoing debate over the validity and value of the two methods.;We first examine the 2008 CDC model for HIV incidence using surveillance data with testing for recent infection. We found that a model-based bootstrap approach corrects for issues of inflated variance and bias in the model. We also show in a simulation study that an additional factor; the number of HIV tests that are motivated by possible exposure or seroconversion sickness, contributes strongly to bias in the model and should be corrected.;The second model we consider is the serial prevalence model first discussed by Ades and Medley in 1994. We adjust the model specification to fit public health STD surveillance rather than testing in antenatal clinics and recast the model in a Bayesian framework in order to allow the inclusion of outside data on testing behavior and assay errors.;Finally, we compare the performance of the two models via a simulation study. The study shows that both models have significant issues with bias and poor coverage for confidence intervals. The serial prevalence model is much less affected by motivated testing than in the STARHS model, but tends to overestimate declines in incidence. Both models mirror incidence trends well, picking up on declines in the underlying incidence rate very rapidly, but significant improvements still need to be made before either method can be considered fully reliable.