The Impact of Network Formation Processes on HIV Spreading Behavior

Introduction: It is well known that the structure of sexual networks affect the spread of HIV, and as such this structure is important to consider in our prevention strategies. However, network data is very hard to collect, especially for sexual networks, and thus is generally incomplete. To make matters worse changes over time (the dynamics of such networks) are largely unknown. To cope with this most simulation models adopt network formation methods to create ‘representative’ networks. While it has been shown that such formation processes can generate structures with similar features, it is largely unknow to which extend the methods employed in the formation process has an effect on the resulting graph structure, and consequently on disease spread dynamics.

Various approaches to network formation exist. A method for generating networks that is being increasingly adopted and is becoming dominant in social sciences is the class of Exponential Random Graph Models (ERGMs). These models generate network structures based on a mechanism that fits a range of local and global level network characteristics for the network as a whole. An alternative approach is suggested in Agent-based Models (ABMs), in which the formation mechanism is based on an individual level decision process from which structure emerges. While these mechanisms clearly differ, it reams unclear if the networks formed, and the disease spread dynamics on such networks, might be different across formation processes.

Method: In this paper we compare the 2 aforementioned network formation approaches; (Separable Temporal) Exponential Random Graph Models ((ST)ERGM) and Agent-based modeling (ABM). As part of a large scale replication effort we have implemented the EpiModel HIV package, originally developed in R, in NetLogo, an Agent-Based Modeling platform widely adopted in Social sciences. Both models behave identically, they employ the same behavioral rules and disease spread dynamics, yet in the NetLogo version the ERGM’s formation process is replaced by an ABM network formation method. This method follows agent level behavioral rules and preferences during the formation process. We compare the networks formed by both processes and measure the effects of the differences in structure on the HIV spread in these systems.

Results: We find that these network formation processes result in networks structures that are in many respects similar, but do show structural differences. We observe that even small differences in the network structure can have a strong impact on HIV spread dynamics. Our result show that network formation process is a component that critically affects the evidence drawn from network models. As such it highlight the importance of accurate network formation processes and the need for detailed network (dynamic) data collection in prevention science.