Estimation of HIV Prevalence at the ZIP Code-Level in Atlanta, Georgia: Bayesian Prediction Modeling Using Passive Surveillance Data and Social Determinants of Disease Spreading. (Preprint)

Abstract

Background: Better information at the ZIP Code-level has the potential to enhance interventions targeting, identify treatment gaps, and optimize resources utilization. Currently there are no methods designed to estimate undiagnosed HIV cases at jurisdictions smaller than counties.Objective: This study aims to predict the number of undiagnosed HIV cases at the ZIP Code-level in Atlanta, Georgia, based on publicly available information. Methods:The CDC reports both passive surveillance (PS) and estimated total (MS) HIV cases for selected counties as part of the Ending of the HIV Epidemic initiative. We employed a Bayesian hierarchical model to: 1) Model MS as random draws from a Poisson distribution with mean equal to the true total HIV cases in the county. 2) A Binomial model for PS arising from the true denominator, with mean P, known as the ascertainment probability. 3) Use a logistic fractional model to allow P to be dependent on socio-economic determinants of HIV extracted from the American Community Survey. These determinants were chosen through a feature selection algorithm. The prediction model was tested out-of-sample on Georgia counties. Finally, we combined zip-code-level covariate data with the posterior predictive distribution of the logit coefficients to predict the mean P at zip-code-level. Final estimates were spatially-smoothed and aggregated to county-level for secondary validations.