Abstract of Distinction: Enriching Agent-Based Models for with Local Data to Support Decision Making on HIV Prevention

Background: Agent-based models (ABMs) are increasingly being adopted to support local health departments’ decision making under uncertainty to optimize prevention efforts. For decision makers such simulation models can be a valuable tool for formulating preventive strategies and monitoring their execution. However, their value strongly depends on the model’s capacity to capture local characteristics of the phenomenon. As such, enriching Agent-Based Models with local data is a vital step towards supporting what we call model-based decision making.

Methods & Materials: We describe the case of creating an ABM to support the Chicago Department of Public Health (CDPH) in their efforts to implement a city wide initiative to end HIV infections as part of ‘Getting to Zero Illinois’. While CDPH’s aim of zero incidence is clear, what is unclear is how to achieve this reduction most efficiently given the characteristics of the Chicago epidemic.

We discuss a complex model of HIV transmission and its prevention that we implemented in NetLogo. We incorporate various local datasets into our ABM on HIV spread among Men-who-have-sex-with-men (MSM). First, as it is known that the structure of sexual networks can strongly affect the spread of HIV, we base the sexual networks in the model on local sexual behaviors captured in RADAR, a Chicago based longitudinal cohort study with sexual network data spanning 3 years and covering more than 1,000 MSM. Second, to accurately describe the local population we use demographic information from CDPH specific to Chicago. Third, to incorporate the local epidemic characteristics, we use CDPH data on the HIV care cascade to infer testing and treatment behaviors. Lastly, actual incidence data collected by CDPH is used to validate model behavior.

Results: Ensuring high fidelity of the model requires validation against local data, including the yearly incidence rates of HIV by age and race/ethnicity. But aligning the model with demographic data, local partner network structure and sexual risk behaviors was not sufficient. Specifically, African American MSM generally have lower risk behaviors than white MSM, yet their incidence rates are higher. To account for racial/ethnic disparities using social determinants, we included a geographic community level measure of sexually transmitted disease as an ecological “leading indicator” variable. This successfully reproduced race/ethnic HIV incidence disparities as it served to account for the syndemic nature of HIV’s contextual effects of access to care, substance abuse, criminal justice involvement, racism and stigma.

Conclusions: Ensuring short-term predictive validity is vital when using models for decision-making. Our model can be used by local health departments to guide decision making based on locally available data. Even when some local data are not available, simulations can readily examine whether the long-term effects of different allocations of prevention efforts are sensitive to these data. If they are, these simulations can guide local health departments to obtain needed data from micro-evaluations.