Addressing Gaps in HIV Preexposure Prophylaxis Care to Reduce Racial Disparities in HIV Incidence in the United States

Samuel M. Jenness; Kevin M. Maloney; Dawn K. Smith; Karen W. Hoover; Steven M. Goodreau; Eli S. Rosenberg; Kevin M. Weiss; Albert Y. Liu; Darcy W. Rao; Patrick S. Sullivan


Am J Epidemiol. 2019;188(4):743-752. 

In This Article


We previously developed a mathematical model for HIV transmission dynamics for MSM in the United States using the EpiModel software platform,[24] a computational toolkit for simulating epidemics over dynamic sexual networks under the statistical framework of temporal exponential random graph models.[25] Our prior applications investigated the sources of HIV racial disparities among MSM in Atlanta and the potential impact of PrEP for MSM across races.[26,27] This study integrated these 2 research streams to develop the model structure, parameterization, and analyses for simulating PrEP stratified by race and represent PrEP care on a continuum framework. Full methodological details are provided in Web Appendices 1–12 (available at

HIV Transmission and Progression

Our model simulates the dynamics of main, casual, and one-time sexual partnerships for non-Hispanic BMSM and WMSM, aged 18–40 years.[26,27] Predictors of partnership formation included partnership type, number of ongoing partnerships, race and age mixing, and sorting by receptive versus insertive sexual position. For main and casual partnerships, we modeled relational dissolution as a constant hazard reflecting their median durations. All network model terms were stratified by race.

MSM progressed through HIV disease in the absence of antiretroviral therapy with evolving HIV viral loads that modified the rate of HIV transmission.[28] After infection, men were assigned to clinical care trajectories controlling rates of HIV diagnosis, initiation of antiretroviral therapy, and HIV viral suppression.[29,30] Antiretroviral therapy was associated with decreased mortality and lower HIV transmissibility.[31] Other factors modifying the HIV acquisition probability included current infection with other sexually transmitted infections,[32] condom use,[33] sexual position,[34] and circumcision status of the insertive partner.[35]

Parameters for network/behavioral features of the model were estimated from 2 studies of HIV disparities between younger BMSM and WMSM in Atlanta, our target population.[22,36] Involvement was a prospective HIV incidence cohort (n = 803), and the MAN Project was a cross-sectional chain-referral sexual network study (n = 314). Venue-time-space sampling was used for both studies to minimize selection biases. Remaining model parameters for the underlying model (e.g., HIV natural history and antiretroviral therapy clinical effects) were assumed to be common across MSM populations and therefore drawn from secondary literature sources. Methods for data analysis and assumptions for model parameterization are described in greater detail in Web Appendices 1–4 and Goodreau et al.[26]

PrEP Continuum

We represented PrEP based on a 5-step continuum: awareness of PrEP, access to healthcare, likelihood of receiving a prescription, effective adherence, and retention in care. Race-stratified probabilities governing transitions across steps were drawn from 2 PrEP demonstration projects.[11,21] Awareness was estimated as 50% for both races, whereas access was 76% for BMSM and 95% for WMSM. Both were fixed attributes assigned at entry into the network.

Prescription probabilities were 63% for BMSM and 73% for WMSM, simulated as a Bernoulli random draw at the point of clinical evaluation for and precondition of initiating PrEP: diagnostic HIV screening. Screening rates were stratified by race based on empirical data (see Web Appendix 7) but were assumed homogeneous otherwise. Consistent with prior models,[32] we simulated the 4 biobehavioral indications for starting PrEP defined in the US Centers for Disease Control and Prevention guidelines:[37] higher-risk sexual behavior in various partnership configurations or a sexually transmitted infection diagnosis within the prior 6 months. Because indications were time-varying, the probability of a PrEP prescription was therefore a joint function of the race-specific probability of receiving a prescription plus current indications at HIV screening.

Effective PrEP adherence in the model represented men taking ≥4 doses per week across follow-up.[11] Proportions meeting this criterion were 60% for BMSM and 93% for WMSM. Taking PrEP at this dosage was been associated with a 98% relative reduction in HIV acquisition risk per sexual act, following Grant et al..[38] MSM who were adherent to PrEP at this level reduced their condom use by 40%.[39] PrEP discontinuation (the converse of retention) rates were based on observed proportions of MSM with indications who had stopped PrEP by week 48 of follow-up (43.8% for BMSM and 18.3% for WMSM). We transformed these proportions into median times to discontinuation (1.1 years and 3.2 years, respectively) assuming a hypergeometric distribution. We simulated this form of spontaneous discontinuation conditional on having ongoing indications, consistent with our data analysis. In addition, men stopped PrEP if they no longer exhibited PrEP indications (evaluated annually for active PrEP users).[37]

Counterfactual Scenarios

To estimate the causal impact of changes to the PrEP continuum for BMSM, we varied the probabilities for each of the 5 steps individually and jointly. The reference scenario with which all intervention counterfactuals were compared was that no MSM (of either race) were on PrEP. While this scenario does not represent a proposed public health strategy, it provides maximum analytical clarity for estimating HIV disparities before and after the introduction of PrEP. Furthermore, we calibrated the model to race-stratified HIV prevalence estimates in 2013, just after the FDA approval of PrEP.[37]

For individual continuum steps, we set the parameters for BMSM to those marginal values observed for WMSM and then higher levels, while holding other BMSM continuum parameters fixed at their observed levels. The WMSM continuum and all other model parameters, including those governing risk behavior and HIV clinical care, were always held fixed across all scenarios; results here are conditional on that assumption. For scenarios modifying parameters in combination, we varied the BMSM parameters on a relative scale. Scenarios in which BMSM parameters were set to 150% of observed values, for example, multiplied each of the empirical estimates by 1.5 (with individual probabilities capped at 1). For our final analysis, we grouped the 5 continuum steps into 2 factor groups—initiation (awareness, access, and prescription) and engagement (adherence and retention)—and then projected outcomes across a spectrum of relative BMSM values in each group.

Calibration, Simulation, and Analysis

With a starting network size of 10,000 MSM (aged 18–40 years), 50% were initialized in each race, a ratio that approximates the distribution for the Atlanta area and provides analytical clarity.[26] We calibrated our model to observed race-specific HIV prevalence at baseline in an Atlanta-based cohort: 43.4% for BMSM and 13.2% for WMSM.[22] Based on prior work on modeling the causes of these disparities,[26] we incorporated the full 95% confidence intervals of estimated rates of anal intercourse and probabilities of condom use for model calibration. We also implemented race-specific parameters simulating condom failure (due to slippage or breakage), consistently higher in BMSM,[40–42] and diagnostic screening for bacterial sexually transmitted infection (increasing the risk of HIV if untreated), often lower for BMSM.[4] Approximate Bayesian computation methods estimated the values of these parameters best fitting the observed prevalence data.[32,43] The calibrated model provided an excellent fit to these targets. We also successfully externally validated this calibration with an "out-of-model" prediction of HIV prevalence by the interaction of race and age (see Web Appendix 12).

Intervention models simulated each scenario over a 10-year time horizon. For each scenario, we simulated the model 250 times and summarized the distribution of results based on median values and 95% credible intervals. Outcomes were race-specific HIV prevalence and incidence per 100 person-years at risk (PYAR), and the hazard ratio comparing incidence with the no-PrEP reference scenario, all at year 10. The percent of infections averted among BMSM compared the cumulative incidence in each intervention scenario with that of the reference scenario. The number needed to treat (NNT) was the number of BMSM person-years on PrEP required to avert 1 new HIV infection for BMSM. Two disparity indices were calculated to compare PrEP impact for BMSM versus WMSM: The absolute disparity was the difference in incidence rates for BMSM and WMSM, and the relative disparity was the ratio of those rates. Finally, we calculated a prevention index as the difference in hazard ratios associated with PrEP uptake for BMSM and WMSM.