Traditional epidemiological models often rely on the assumption of "homogeneous mixing"—the idea that any individual in a population has an equal probability of contacting any other. While computationally convenient, this simplification fails to capture the reality of diseases like HIV, where transmission is highly dependent on specific social and sexual networks. Even more advanced "static network" models, which treat the population as a fixed graph, struggle with long-term epidemics where the network itself is in a state of constant flux.
In the paper "An SIR Graph Growth Model for the Epidemics of Communicable Diseases," the authors propose a novel graph-valued stochastic process. This model treats the social network not as a static backdrop, but as a dynamic entity that grows and evolves alongside the spread of the disease.
The Model: Graph-Valued Stochastic Processes
The core of the proposed model is a Susceptible-Infected-Removed (SIR) process where the underlying graph is generated through a Poisson process. Unlike standard models, this approach incorporates a "preferential attachment" mechanism, where individuals who already have many contacts are more likely to form new ones—a common feature of real-world social structures.
Each vertex (individual) in the graph is "marked" with specific covariates, such as gender, sexual orientation, and serological status. These marks are not just descriptive; they actively influence the rates of contact, infection, and eventual detection (removal). This allows the model to simulate heterogeneous populations where different subgroups exhibit different transmission dynamics, providing a more realistic representation of how a virus moves through a community over decades.
Statistical Inference via ABC
One of the primary challenges in network epidemiology is the "missing data" problem. Researchers rarely observe the entire social network; instead, they see an "observable subgraph" consisting of diagnosed individuals and their traced contacts. Calculating the exact likelihood of such a partial graph is often mathematically intractable.
To overcome this, the authors utilize Approximate Bayesian Computation (ABC). This method involves simulating thousands of potential epidemics based on different parameter sets. To determine which simulations are "close" to the observed data, the authors employ a sophisticated graph-matching algorithm based on quadratic convex relaxation. By comparing the structural topology—such as cluster sizes and connectivity—of simulated networks against the real-world data, the model can infer the underlying parameters of the hidden epidemic without needing a closed-form likelihood function.
Application: The Cuban HIV Epidemic
The model was validated using a unique dataset from Cuba, covering 5,389 individuals diagnosed with HIV between 1986 and 2006. Cuba’s intensive contact-tracing system provided a rich "observable network" of 4,073 sexual contacts, making it an ideal case study for this methodology.
Key Results:
- Detection Trends: The model successfully captured a shift in the Cuban epidemic's dynamics. While contact tracing was the primary driver of diagnosis in the early years, spontaneous or "random" detections eventually outpaced it as the epidemic expanded into different sub-populations.
- Structural Accuracy: The model showed remarkable precision in reproducing the network's physical structure. For the 1998–1999 period, the model predicted a Largest Component (LC) size of 1,066 individuals, which closely aligned with the actual observed LC of 1,133.
- Parameter Recovery: In toy simulations, the ABC method proved capable of narrowing down true parameters (like infection probability) even when dealing with high-dimensional missing data.
Insights and Future Directions
The paper highlights that the epidemic and the social network exist in a feedback loop: the disease spreads through the network, but the network's structure is simultaneously reshaped by the disease as individuals are diagnosed and "removed" from the active transmission pool. By focusing on the shape and topology of the network rather than just raw infection counts, the authors provide a more robust framework for understanding long-term disease dynamics.
While the computational cost of graph matching remains a hurdle for scaling to populations of millions, this research provides a vital foundation for using dynamic graph models in public health. Future work may involve "online" graph matching or parallelized ABC to handle the massive datasets generated by modern digital contact-tracing applications.
Summary updated