Modelling the dispersion of SARS-CoV-2 on a dynamic network graph

Patrick Bryant, Arne ElofssonmedRxiv 2020 · DOI

Abstract

Traditional epidemiological models often rely on the basic reproductive number ($R_0$), which assumes a uniform distribution of social contacts. However, social networks exhibit significant variation in connectivity, with highly connected individuals acting as "superspreaders." This study models the dispersion of SARS-CoV-2 on a dynamic preferential attachment graph that evolves in response to observed mobility changes. By incorporating a serial interval distribution and age-specific infection fatality rates, we demonstrate that the social network structure plays a critical role in epidemic development. Our results show that saturation can be reached much earlier on a preferential attachment graph than on a random graph, as the infection of highly connected hubs rapidly reduces the network's overall connectivity. This highlights the importance of accounting for non-uniform social structures in public health modeling.

Introduction

Most standard epidemiological models, such as the classic SIR model, assume a "well-mixed" population where every individual is equally likely to infect any other. This assumption is encapsulated in $R_0$, an average value that overlooks the heterogeneity of human social interactions. In reality, social networks are composed of individuals with vastly different numbers of contacts. "Hubs" with many connections drive the initial surge of an outbreak and act as superspreaders.

This paper addresses the limitation of average-based models by investigating how dynamic network structures and mobility changes influence the trajectory of COVID-19. The authors argue that once these highly connected individuals become immune, the network’s overall "connectedness" collapses, making it much harder for the virus to spread even if a large portion of the population remains susceptible. This phenomenon may explain observed trends in early hotspots like Spain and New York City.

Methodology

The researchers utilized a dynamic preferential attachment graph based on the Barabási–Albert model. In this framework, new nodes are more likely to connect to existing nodes that already have many links, naturally creating the hubs missing from standard models.

Preferential attachment graphs with varying density parameters (m=1, 2, 3).

To simulate real-world conditions, the network was made dynamic by rearranging links based on Google Community Mobility Reports. Increased mobility led to a higher probability of new connections, while decreased mobility (e.g., staying at home) reduced it. The model also sampled a serial interval distribution to determine transmission timing and used age-specific infection fatality rates (IFRs) for four distinct age groups.

The model's performance was compared against a random network (Erdős–Rényi model) baseline to quantify the impact of network topology on epidemic forecasting.

A random graph (p=0.06) used as a baseline comparison for the preferential attachment model.

Results and Discussion

The Hub Effect and Saturation

In a preferential attachment graph, the virus naturally finds and infects hubs early in the outbreak. Because these hubs are removed from the susceptible pool (through recovery or death), the epidemic reaches a saturation point much faster than in a random network. The model suggests that saturation—where the spread slows significantly—can occur at much lower infection percentages than traditional herd immunity estimates (often 60-80%). For instance, in Spain, the best fit showed spread slowing at just 12.5% infection.

Network Density and Geography

The parameter $m$, representing the number of links introduced with each new node, reflects social density. The study found that New York City (NYC) was best modeled with $m=2$ (higher connectivity), while Spain was best modeled with $m=1$. This suggests that denser urban environments require different modeling parameters than broader national populations.

Model Validation and Timing

The preferential attachment model aligned closely with real-world data. In NYC, peak deaths were modeled in mid-April, matching observations. Conversely, a random graph with the same parameters predicted a peak in mid-June, illustrating the danger of ignoring network structure. The best fits for both regions were achieved by combining dynamic network changes with a 33% reduction in infection probability, representing non-pharmaceutical interventions like masking.

Conclusion

This study provides a novel method for modelling epidemics by linking dynamic network structures to observed mobility changes. The social network structure is a crucial, yet often overlooked, factor in epidemic development. By accounting for the "superspreader" effect and the rapid collapse of network connectivity as hubs become immune, we can better understand why some regions reach saturation points earlier than predicted by traditional models. While limited by computational scaling and the use of synthetic networks, this approach offers a more nuanced framework for public health planning.

Summary updated

Abstract

Abstract Background When modelling the dispersion of an epidemic using R 0 , one only considers the average number of individuals each infected individual will infect. However, we know from extensive studies of social networks that there is significant variation in the number of connections and thus social contacts each individual has. Individuals with more social contacts are more likely to attract and spread infection. These individuals are likely the drivers of the epidemic, so-called superspreaders. When many superspreaders are immune, it becomes more difficult for the disease to spread, as the connectedness of the social network dramatically decreases. If one assumes all individuals being equally connected and thus as likely to spread disease as in a SIR model, this is not true. Methods To account for the impact of social network structure on epidemic development, we model the dispersion of SARS-CoV-2 on a dynamic preferential attachment graph which changes appearance proportional to observed mobility changes. We sample a serial interval distribution that determines the probability of dispersion for all infected nodes each day. We model the dispersion in different age groups using age-specific infection fatality rates. We vary the infection probabilities in different age groups and analyse the outcome. Results The impact of movement on network dynamics plays a crucial role in the spread of infections. We find that higher movement results in higher spread due to an increased probability of new connections being made within a social network. We show that saturation in the dispersion can be reached much earlier on a preferential attachment graph compared to spread on a random graph, which is more similar to estimations using R 0 . Conclusions We provide a novel method for modelling epidemics by using a dynamic network structure related to observed mobility changes. The social network structure plays a crucial role in epidemic development, something that is often overlooked.
BibTeX
@misc{bryant2020modelling,
  author = {Patrick Bryant and Arne Elofsson},
  title = {{Modelling the dispersion of SARS-CoV-2 on a dynamic network graph}},
  year = {2020},
  doi = {10.1101/2020.10.19.20215046},
  howpublished = {medRxiv},
}
Figure