We sought to conduct the pilot in health departments that varied by HIV prevalence, capacity, resources, and geography. On the basis of those criteria, we invited participation by the Chicago Department of Public Health (Chicago), the Nebraska Department of Health and Human Services (Nebraska), and the Alabama Department of Public Health (Alabama). Chicago and Philadelphia represented jurisdictions with higher HIV prevalence, higher capacity, and greater resources, whereas Nebraska and Alabama represented lower HIV prevalence jurisdictions with fewer resources.

The model was piloted sequentially in Chicago, Nebraska, and Alabama so that lessons learned from one deployment could be incorporated into the next. The process began at each site with an introductory call and was followed by the provision of basic information about modeling and resource allocation as well as a codebook for HIV-RAMP data inputs. We then introduced the resource allocation model and demonstrated how to develop inputs, run the model, interpret the results, and conduct, interpret, and understand sensitivity analyses.

Using Microsoft Excel, Visual Basic for Applications, and Solver add-ins, we developed a linear programming model combined with a Bernoulli process model to optimize HIV prevention resource allocation for health departments participating in the pilot (see the Technical Appendix, Supplemental Digital Content, available at: http://links.lww.com/JPHMP/A167). We considered HIV prevention interventions that were aligned with the principles of high-impact prevention, were recommended by CDC or other federal agencies, and for which sufficient efficacy data existed. These included (a) testing in clinical settings, (b) testing in nonclinical settings, (c) partner services, (d) continuum-of-care–related interventions designed to improve linkage to care, retention in care, and adherence to antiretroviral therapy (ART), and (e) behavioral interventions for HIV-positive and HIV-negative persons. Testing in nonclinical settings and behavioral interventions were targeted to specific risk populations such as MSM, IDUs, and sexually active heterosexuals (HETs). Nontargeted interventions included testing in clinical settings, partner services, linkage to care, retention in care, and adherence to ART. We excluded structural interventions such as condom distribution, syringe exchange, and social media campaigns, because we could not identify sufficient data for use in a Bernoulli process model, described in the following text, to estimate a per person reduction in the risk of HIV transmission or acquisition.

We used a Bernoulli process model to estimate the annual risk of HIV transmission or acquisition, which required behavioral data specific to each risk population on the average number of partners per year, the number of sex acts per partner, condom use, estimated HIV prevalence among partners, and partnership overlap. In addition, the model utilized estimated per act HIV transmission probabilities for each type of sex act. For the IDU population, we also considered transmission or acquisition by needle sharing. We estimated the effects of HIV prevention interventions by reducing per act or per partnership transmission probabilities or increasing the proportion of condom-protected sex acts. We estimated an intervention’s effect on annual HIV transmission or acquisition risk as the difference in the Bernoulli process model’s calculation of annual transmission risk with and without the intervention’s effect. Then, we determined the cost per case of HIV infection prevented for each intervention by dividing the cost of the intervention by the intervention effect. Because continuum-of-care parameters—such as the proportion of undiagnosed HIV-infected persons among those living with HIV infection—varied by pilot site, the value and rankings of the same continuum-related intervention could differ across the pilot sites.

We applied various periods of duration to the interventions. We assumed reductions in risky behavior associated with a new diagnosis of HIV infection, including one delivered through partner services, lasted for 5 years. In a meta-analysis by Marks et al,¹⁵ of high-risk sexual behavior in persons aware and unaware of their HIV infections in the United States, the prevalence of unprotected anal or vaginal intercourse was similar among HIV-positive aware men regardless of the length of time between when they were surveyed and when they became aware of their infection (up to 8 years). Because studies evaluating the efficacy of behavioral interventions typically followed participants for 12 months or less^16–20 and few studies on care and treatment-related interventions followed the participants for more than 2 years,^21–25 we assumed a 1-year duration for behavioral interventions and a 2-year duration for continuum-of-care interventions, such as retention in care and adherence to treatment. Further discussion of our assumptions about duration can be found in the Technical Appendix (see Supplemental Digital Content, available at: http://links.lww.com/JPHMP/A167). After accounting for duration, the site-specific cost per case of HIV infection prevented was used as an input to the optimization model.

For the optimization model, we also set the maximum reach, or the maximum proportion of the eligible population that could be reached by the intervention. The proportion reflected assumptions about the fraction of the eligible population that would avail itself of the intervention during the course of a year and intervention scalability. We also considered the proportion served by HIV prevention interventions funded by the Philadelphia health department during a 1-year period. We set the maximum reach at 20% for continuum-of-care interventions and behavioral interventions for HIV-infected persons, 10% for testing and behavioral interventions for uninfected persons, and 5% for partner services. Given lack of data from the 3 pilot sites, we set the maximum reach consistently for each intervention across the sites.

The objective of the optimization model was to maximize the number of new infections prevented for a given health department budget, achieved through the following steps:

The model inputs included epidemiologic, behavioral, budgetary, and programmatic data. The Bernoulli process model used behavioral data and intervention efficacy to determine annual transmission and acquisition risk, with and without each intervention. The optimization model used the budget amount, intervention costs, and each health department’s epidemic profile, including the maximum reach of each intervention.

Pilot site data included their total CDC HIV prevention budget, site-specific HIV prevalence and annual incidence by risk group, and the proportion of those infected or diagnosed with HIV infection who achieved steps along the continuum of care: diagnosed, linked to care, retained in care, and adherent to treatment (Table 1).

The data CDC provided included behavioral data and estimates of intervention costs and efficacy. The behavioral data included per act HIV transmission or acquisition probabilities by sex act and contaminated needle sharing, the annual number of partners and injections, the number of sex acts per partner, the proportion of sex acts protected by condoms, the proportion of needle/syringe shared among users, and the proportion of infected persons having more than 1 infected partner at a time. We derived parameter estimates from the National Survey of Sexual Health and Behavior^26–28 and the National Survey of Family Growth.^29–31 We used expert opinion when relevant data were unavailable.

Using program expenditures reported by health departments and costs from trials and published studies, we estimated the cost to a health department to deliver an intervention. All costs were converted to 2009 US dollars using the medical care component of the Consumer Price Index.³² Philadelphia supplied its own program costs and some program efficacy data based on health department–specific budget and outcome data.

We conducted the analysis from a health department perspective. Model inputs for duration of effect, maximum reach, and behavior change following a new diagnosis were evaluated in sensitivity analyses. Those methods and results are reported in the Technical Appendix (see Supplemental Digital Content, available at: http://links.lww.com/JPHMP/A167).

The model generated the optimal allocation of funds, the site-specific cost per case of HIV infection prevented rankings, the total number of persons served, the expected number of HIV cases prevented, and the total HIV life time treatment cost saved if the optimal allocation was realized for each site. Total HIV life time treatment costs saved typically are not incurred by health departments. However, this measure quantifies the health department’s prevention efforts and presents the overall benefit of the intervention.

We estimated the cost and efficacy of interventions compared with the baseline/status quo that would have occurred without the intervention. This baseline/status quo reflects the impact of previous HIV prevention funding. We measured the number of HIV cases prevented compared with baseline/status quo for each intervention.

We focused on 4 evaluation research questions:

We evaluated the health departments’ experience during site visits and through interviews conducted by JSI. Although JSI provided the technical assistance and evaluated the health departments’ experience using the model, different teams within JSI were assigned to conduct those 2 tasks to avoid conflict of interest.

The model output indicated funding for HIV testing and continuum-of-care interventions in all 4 sites but not for behavioral interventions for HIV-positive or HIV-negative persons (Table 2). As a result, we combined behavioral interventions for HIV-positive and HIV-negative persons in Table 2. In 3 of the 4 health departments, results suggested that the majority of funds be allocated to continuum-of-care interventions, with less than half allocated to HIV testing. The exception was Philadelphia, where 60% of the budget was allocated to testing.

The results suggested that as much as 50% of the prevention budget be allocated to specific continuum-of-care interventions. The site-specific allocation depended upon the proportion of HIV-infected people who already had achieved each step of the continuum, as well as the size of the site’s CDC budget.

The largest proportion of overall testing funds was allocated to testing in clinical settings. Funds were also allocated to HIV testing of specific risk groups outside of clinical settings based on the relative size of the risk group’s infected but undiagnosed population. The model allocated funding for HIV testing through partner services only in Chicago. It did not allocate funding to HIV testing of HETs in any nonclinical settings.

The cost-effectiveness ranking of HIV prevention interventions varied little across the 4 health departments. The most cost-effective intervention for all sites was HIV testing for MSM in nonclinical settings, whereas behavioral interventions were the least cost-effective in all sites. Behavioral interventions for HIV-positive persons ranked higher than those for HIV-negative persons, and behavioral interventions for MSM were more cost-effective than those for IDUs and HETs. The cost-effectiveness of the targeted interventions reflects the underlying differences in the probability of HIV transmission or acquisition by risk group.

The expected number of persons served with all interventions ranged from 248 in Nebraska to 2723 in Chicago (Table 2). The expected number of HIV cases prevented, given the optimal allocation of 1 year of CDC prevention funds, ranged from 33 in Nebraska to 430 in Chicago. The average cost per case prevented ranged from $25 350 in Alabama to $48 765 in Philadelphia, and life time HIV treatment costs saved from 1 year of prevention funding ranged from $12 million in Nebraska to $158 million in Chicago.

Health departments required skilled staff to use Microsoft Excel and to collect the budgetary, surveillance, and program data inputs needed to implement the model. Four to 7 health department staff members participated in the project at each site. Their tasks included developing model inputs, running the model, and reviewing and presenting the results. Technical assistance provided by JSI ranged from 84 to 110 days across the 3 pilot sites. The sites varied in their ability to assemble the necessary data inputs for continuum-of-care outcomes, new diagnoses of HIV infection, and the total number of persons living with HIV infection in their jurisdictions by transmission risk group (ie, MSM, IDUs, HETs).

Once trained and with the best available inputs supplied to the model, health department staff found the quick run time of the model and its ease of use and interpretation to be most useful. Participants reported that the model had the potential to guide them on allocating their budget efficiently to maximize the number of HIV infections prevented. The average cost per case of HIV infection prevented by an intervention was one of the most useful outputs. In some instances, the pilot project enhanced collaborations between health department surveillance and program staff. The model results provided rigorous and quantitative evidence for decision making about how to distribute HIV prevention resources among programs and populations. The results were useful for justifying funding allocations to program managers, people living with HIV infection, and HIV prevention advocacy groups. On the basis of its results, Philadelphia planned to reduce funds for behavioral interventions and interventions targeting HETs. Alabama planned to increase funds for retention in care. Alabama and Chicago planned to reduce funds for behavioral interventions, particularly for HIV-negative individuals.

Participants identified the greatest limitations of HIV-RAMP as the exclusion of structural interventions, the inability to incorporate non-CDC funds for HIV prevention and treatment interventions, and the lack of integrated interventions that encompassed several steps of the continuum of care. In addition, some sites found that they could not respond to the model recommendations in a timely manner if the model results became available only after multiyear contracts were established.

Participants recommended that future model users should be clear about the model’s purpose and limitations, data inputs required, and inclusion and exclusion criteria for interventions. The participants also stated it would be important for future users to realize that models offer guidance on relative allocations of funds rather than on precise amounts to be allocated and that models may not incorporate important factors such as feasibility and equity.