In five of six real-world group networks, a less tidy pattern of mixing was associated with larger average outbreaks—but with less uncertainty about which person started them. The exception was a network of primary-school contacts. The finding is a warning against assuming that epidemics behave the same way in every network: who shares groups with whom can matter, even when the number and sizes of those groups stay fixed.

The study concerns hypergraphs, a way to represent interactions involving more than two people at a time. In a conventional network, an edge joins a pair. A hyperedge can join a whole discussion group, a set of people reviewing the same kind of restaurant in one month, or a cluster of students in close contact. That distinction matters for contagion: one infected person in a group may expose several susceptible people in one step.

The researchers, Jelmer Stolwijk and colleagues at Delft University of Technology, examined how one particular mixing pattern relates to spread under the Susceptible–Infected Contact Process, or SICP. Their central result is not that one universal network arrangement always drives outbreaks. It is that an association visible in several empirical networks largely disappears in randomized synthetic ones. Network structure is not background scenery; it can shape what a spreading process does.

The Science

The feature they varied is assortativity by the neighborhood coefficient, abbreviated NC. A node’s NC is the average size of the hyperedges it belongs to. Someone whose interactions tend to happen in large groups has a higher NC than someone whose groups are usually small. NC assortativity asks whether nodes with similar values tend to share the same hyperedges.

To change assortativity without changing each node’s basic participation profile, the researchers used degree-preserving rewiring. They selected two hyperedges of the same size and swapped one member between them. The swap keeps both group sizes intact and preserves the number of hyperedges of each size that contain each node. In effect, it changes who is grouped with whom while retaining a detailed record of how often each person belongs to groups of each size.

They also defined a computationally simple measure of hypergraph assortativity. It treats the overall variation in a node property across all node–group memberships as the sum of variation within groups and variation between group averages. The measure reports the fraction of total variation explained by differences between hyperedges: high values mean members of each group tend to be alike on the chosen property. This is an analysis-of-variance idea, adapted to group networks.

For spreading, the SICP starts with one infected or informed node. At each discrete time step, every infected node chooses one of its hyperedges uniformly at random and attempts to infect each susceptible member of that group, independently with probability . The simulation ends after  steps. The team averaged outcomes over 1,000 independent runs for every possible starting node, then compared both mean final prevalence and the variation across seeds.

The empirical datasets represent very different kinds of groups: MathOverflow participants answering algebra or geometry questions; Yelp users reviewing restaurant or bar categories in a month; Amazon users reviewing blues music; and students in a French primary school whose close contacts were measured at the same time. The datasets ranged from 242 nodes in the school-contact network to 1,234 in Bars-Rev. The team focused its rewiring experiment on three-person hyperedges, in part because those were more numerous than larger groups in several datasets.

What They Found

In five of the six real-world networks, lower NC assortativity tended to go with higher average prevalence. Put another way, when nodes with similar average group sizes were less tightly clustered together, the infection process tended to reach more people on average. The relationship was not reported for contact-primary-school. The paper describes the shifts in prevalence as generally small, and the supplied text does not provide numerical values from the plotted curves; the result is a consistent direction in most empirical cases, not a quantified promise of a large change.

The authors offer a mechanism. A node with a high NC tends to participate in larger groups. If infected, it may choose one of those groups and have more susceptible people available to infect. High-NC nodes can therefore be both more likely to be reached through large groups and more capable of passing infection onward. Under high assortativity, they tend to share groups with other high-NC nodes. The researchers argue that these contacts can be redundant: the infection connects people who already have comparatively strong access to large groups, rather than opening as many routes into other parts of the network.

That explanation is plausible, but it should not be mistaken for a universal rule that high assortativity suppresses epidemics. The authors emphasize that the observed effect depends on the broader topology. They generated random synthetic hypergraphs with matching numbers of nodes and hyperedges of each size, then compared those with rewired versions. In most such synthetic networks, including the example shown for Restaurants-Rev, the negative relationship between assortativity and prevalence was largely absent.

The distinction points to a hidden ingredient in the real datasets: heterogeneity. The five networks showing the trend had non-normal NC distributions and included nodes with high NC but relatively few total hyperedges. A person who belongs to only a few groups may nonetheless be part of unusually large groups. That combination can create routes and bottlenecks that a random model with the same broad counts does not reproduce. The school-contact network, by contrast, and the random examples had NC distributions closer to normal, according to the paper.

A second result concerns the identity of the first infected person. As NC assortativity rose, the standard deviation of final prevalence across possible seed nodes generally increased, especially in those same five empirical networks. In a strongly assortative arrangement, a low-NC seed may mostly reach other low-NC nodes and generate modest spread; a high-NC seed may reach a chain of high-NC nodes with access to large groups. The outbreak’s eventual size becomes more sensitive to where it begins.

This does not mean that high assortativity makes outcomes unpredictable in every practical sense, or that the study measured uncertainty in real outbreaks. The reported variation is a model comparison across possible single seeds under the study’s fixed infection probability and time horizon. Still, it highlights a policy-relevant distinction: two networks can have similar group counts and the same number of group memberships per person, yet differ in how strongly the starting location affects the modeled outcome.

Why This Changes Things

Many public-health models and information-spread analyses begin with pairwise contacts. That can be a useful simplification, but it may miss the fact that a single encounter can expose several people together. Hypergraphs preserve this group structure directly. A lecture, meeting, online forum, or shared workplace shift is not simply a collection of independent pairs; the group itself can create a common exposure event.

The paper adds a practical tool to this field: a simple assortativity measure and a rewiring procedure that can adjust it while keeping each node’s group-size-specific participation counts fixed. This helps isolate one structural feature from others. If a modeled outcome shifts after such rewiring, the comparison is less likely to be explained merely by changing how many groups each person attends or by changing group sizes.

The intriguing implication is that interventions may need to consider not only how many contacts people have, but how those contacts are bundled. If high-NC people repeatedly share groups, interventions aimed at a few central venues or groups could have different effects than interventions that connect those people to a wider range of network regions. Conversely, identifying high-NC individuals alone may not be enough: their role depends on whether they connect otherwise separate parts of the system or mainly encounter one another.

These are implications for modeling and measurement, not ready-made instructions for schools, workplaces, or health agencies. The study simulates an abstract infection process; it does not establish that a particular real pathogen, misinformation campaign, or intervention would respond in the same way. The SICP assumes a fixed transmission probability and a rule in which infected nodes select one incident group uniformly. Real behaviors, repeated exposures, timing, and protective measures can all change transmission.

The contrast between empirical and random networks is especially valuable. It cautions against treating a synthetic network that matches a few summary statistics as a faithful stand-in for a real contact system. A model may preserve group counts and still miss the combination of unusual group sizes, uneven participation, and correlated memberships that shapes spread. For anyone using networks to plan epidemic response or information campaigns, that is a reason to validate more than the network’s average degree or average group size.

What’s Next

The authors leave several important questions open. Their experiments rewired three-person groups, even though some datasets contain much larger hyperedges. Larger groups could matter more because one infected member has more potential recipients, but the paper does not establish that rewiring them would produce a larger effect. Testing different hyperedge sizes separately—and combinations of sizes—would show whether the result is specific to the three-person interactions studied.

A second challenge is explaining exactly which topological ingredients make the trend appear. The authors point to non-normal NC distributions and nodes with high NC but low total hyperdegree, but they do not isolate those features experimentally. Synthetic hypergraphs that vary these properties systematically could distinguish whether the decisive factor is the shape of the NC distribution, the association between NC and total group participation, or some interaction among them.

The work also presents a directional finding without numerical effect sizes in the text supplied here. The plotted results indicate trends, while the paper describes changes as generally small. Replication with confidence intervals, sensitivity tests across infection probabilities and durations, and alternative spreading rules would help establish how robust and practically consequential the patterns are.

Even with those limits, the study makes a useful point: group structure can change not only how much a contagion spreads, but how much the outcome depends on its first foothold. Better maps of real group interactions could help researchers identify where simple pairwise models miss important pathways. The next step is to learn which features of those maps matter most—and when changing the arrangement of groups can genuinely make a system more resilient.