Correlated Superspreading in Epidemics

• Correlated superspreading is a phenomenon where a small fraction of cases drive epidemic clusters through statistically dependent transmission events across individuals and shared environments.
• It manifests through mechanisms such as autocorrelated individual infectiousness, environmental synchrony in enclosed spaces, and assortative mixing in network structures.
• Empirical analyses and models show that targeting high-risk settings and key network cores can markedly reduce outbreak propagation.

Correlated superspreading denotes superspreading that is structured rather than i.i.d.: transmission is concentrated in a small minority of infected individuals or events, but that concentration is also statistically linked across infectors and infectees, shared environments, contact networks, or repeated temporal contexts. In the standard epidemiological sense, superspreading implies a fat-tailed distribution of infectiousness or offspring number, so that a small fraction of infected individuals causes a large majority of new infections (Pozderac et al., 2020). Correlated superspreading adds dependence to that heterogeneity. In different parts of the literature, this dependence is represented as autocorrelation of individual reproductive numbers along chains of transmission, shared exposure to a common aerosol field in enclosed spaces, assortative placement of high-transmission individuals in dense network cores, or explicit correlation between “risk” and “spread” in directed contact graphs (Leonardi et al., 18 Sep 2025).

1. Conceptual scope

Across epidemic theory, network epidemiology, and environmental transmission models, correlated superspreading is not a single formalism but a family of related mechanisms. In one line of work, it means that high-transmission infectors tend to generate high-transmission infectees, so that superspreading propagates along transmission chains rather than being redrawn independently at each infection (Leonardi et al., 18 Sep 2025). In another, many people in the same enclosed space are jointly exposed to a shared, time-evolving aerosol concentration that crosses a common infective threshold, producing spatially and temporally clustered infections (Kolinski et al., 2020). Network-based formulations instead emphasize assortativity, high betweenness bridges, dense $k$-cores, or correlations between degree and edge infection rate, so that superspreading recurs in the same structural regions of the contact graph (Mizutaka et al., 2021).

This broader view departs from homogeneous or mass-action intuitions. Kroy’s synthesis describes epidemic spreading as a fundamentally heterogeneous and erratic process with fat tails and rare but consequential bursts, closer in some respects to earthquakes, hurricanes, traffic jams, and stock crashes than to diffusive mixing in a test tube (Kroy, 2022). A common misconception is therefore that superspreading is exhausted by a single overdispersed offspring distribution. The literature summarized here shows instead that the same overdispersion can have very different population-level consequences depending on whether it is independent, environmentally synchronized, temporally correlated, or embedded in assortative network structures.

2. Individual infectiousness, offspring distributions, and correlation at the host level

A standard quantitative starting point is individual infectiousness $\beta$, defined as the number of secondary infections caused per infected individual per day. If, for an individual with infectiousness $\beta$, daily infections are Poisson, then the population-level distribution of daily secondary infections is a Poisson mixture,

$$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$

with moments

$$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$

When $p(\beta)$ is gamma, $P(n)$ is negative binomial with dispersion parameter $k=\mu_\beta^2/\sigma_\beta^2$, so small $k$ corresponds to a fat tail and strong superspreading (Pozderac et al., 2020).

County-level early-pandemic data from the United States yielded $\mu_\beta = 0.18$ cases/day and $\beta$0 cases/day, giving $\beta$1 and $\beta$2. Under the gamma assumption, the inferred Lorenz curve implies that the top $\beta$3 of individuals generated about $\beta$4 of infections, the top $\beta$5 generated $\beta$6, and the top $\beta$7 generated $\beta$8; equivalently, only about $\beta$9 of cases arose from the $\beta$0 of infected individuals with the lowest infection rates (Pozderac et al., 2020). These values quantify strong superspreading, but by themselves they do not specify whether high infectiousness is independent across persons or correlated along transmission paths.

A second host-level correlation enters when infectiousness and susceptibility are themselves linked. In the heterogeneous herd-immunity model of (Oz et al., 2020), each individual $\beta$1 has susceptibility $\beta$2 and infectiousness $\beta$3, with transmission probability

$\beta$4

The key object is

$\beta$5

The fully correlated case takes $\beta$6, so highly susceptible individuals are also highly infectious. This is a direct mathematical version of correlated superspreading: the same individuals are more likely to become infected early and to generate many onward infections once infected.

3. Environmental, spatial, and temporal correlation

A distinct formulation treats superspreading as environmental amplification. In a well-mixed room of volume $\beta$7, with aerosolized virion emission rate $\beta$8, destabilization rate $\beta$9, and optional filtration $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$0, the aerosol concentration satisfies

$$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$1

so that with zero initial concentration,

$$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$2

Individual exposure over duration $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$3 is

$$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$4

Across five poorly ventilated enclosed-space events with a single spreader and resting respiratory activity, estimated $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$5 values all lay between 50 and 100 virions, suggesting a possible unique aerosol minimum infective dose of $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$6. Extending the analysis to 20 superspreading events, almost all aerosol-compatible events again clustered around $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$7 virions (Kolinski et al., 2020). In this framework, infections are correlated because many occupants inhale the same time-dependent aerosol cloud; once the shared cumulative dose exceeds the effective threshold, many infections occur together.

This environmental mechanism aligns with broader empirical regularities. A synthesis focused on SARS-CoV-2 emphasizes that many explosive superspreading events occurred in indoor settings such as long-term care facilities, prisons, meat-packing plants, fish factories, cruise ships, family gatherings, parties, and night clubs, and cites a Japanese cluster analysis in which closed environments promoted superspreading events with an odds ratio of approximately $$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$8 relative to open spaces (Althouse et al., 2020). The repeated appearance of these settings indicates that superspreading is correlated with closed environments, poor ventilation, crowding, and long durations of exposure, rather than being distributed uniformly across all contacts.

Temporal correlation, however, does not always amplify spread. In temporal-network SI models with bursty contact processes, positive correlation between consecutive interevent times is quantified by a memory coefficient

$$P(n)=\int_0^\infty d\beta\,\frac{e^{-\beta}\beta^n}{n!}p(\beta),$$9

For two-step deterministic SI, the mean transmission time becomes

$$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$0

and positive $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$1 slows spreading on Bethe lattices and random graphs by increasing the shortest transmission time from infected nodes to their susceptible neighbors (Hiraoka et al., 2018). This contrast is important: some temporal correlations synchronize exposure and create clusters, while others delay propagation along links.

4. Network assortativity, superspreading cores, and directed risk–spread structure

Network models make correlated superspreading a structural property of subgraphs rather than isolated individuals. In synergistic SIS dynamics on a bimodal network, a susceptible node with $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$2 infected neighbors is infected at rate

$$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$3

so transmission is superlinear in the number of infected neighbors when $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$4. Degree correlations are encoded by an assortativity coefficient $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$5. Positive degree correlation lowers epidemic thresholds, negative degree correlation raises them, and with strong synergy and strong positive correlation the mean-field theory predicts two discontinuous jumps in steady-state prevalence, corresponding to explosive invasion first of a high-degree community and later of a low-degree community (Mizutaka et al., 2021). This is correlated superspreading in the strict network sense: high-degree nodes are clustered with one another, and synergy makes dense infected neighborhoods disproportionately dangerous.

Mobility-based contact-network reconstruction during COVID-19 gives a directly empirical version of the same idea. Using approximately 450 million GPS data points per day across Latin America, the network analysis of (Serafino et al., 2021) found that lockdowns reduced mobility by about $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$6 and caused the transmission network to disintegrate by about $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$7, yet the epidemic persisted because superspreading $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$8-core structures remained. The outer $$\mu_n=\mu_\beta,\qquad \sigma_n^2=\mu_\beta+\sigma_\beta^2.$$9-kshell dropped by about $p(\beta)$0, closely tracking the giant connected component, whereas the inner $p(\beta)$1-kcore dropped by only about $p(\beta)$2 and later increased again. These persistent cores localized at hospitals, warehouses, business buildings, and large condominiums. Removing high-betweenness “weak links” connecting the large $p(\beta)$3-cores dismantled the giant component with only $p(\beta)$4–$p(\beta)$5 node removal, much more efficiently than hub or $p(\beta)$6-shell targeting (Serafino et al., 2021).

Directed-graph theory sharpens the same point analytically. If $p(\beta)$7 denotes the number of incoming transmission edges (“risk”) and $p(\beta)$8 the number of outgoing transmission edges (“spread”), then for a random directed graph with joint distribution $p(\beta)$9,

$P(n)$0

Positive correlation between $P(n)$1 and $P(n)$2 therefore raises $P(n)$3 directly. The same framework yields forward and backward versions of the friendship paradox: forward edges tend to lead to individuals with high risk, while backward edges lead to individuals with high spread, implying that backward tracing is particularly effective for locating superspreading sources and preventing further cascades (Allard et al., 2020).

5. Thresholds, epidemic amplification, and herd immunity under correlation

The most explicit epidemic-threshold theory for correlated superspreading models transmission potential itself as autocorrelated along infector–infectee chains. If $P(n)$4 is the individual reproductive number of infection $P(n)$5, then

$P(n)$6

where $P(n)$7 denotes the parent infection and $P(n)$8 is an autocorrelation parameter. Because high-$P(n)$9 individuals are more likely to become parents, the critical threshold is not set solely by $k=\mu_\beta^2/\sigma_\beta^2$0. The resulting condition for persistence is

$k=\mu_\beta^2/\sigma_\beta^2$1

so chains of superspreading can sustain epidemics even when the average transmission rate in the host population is below one (Leonardi et al., 18 Sep 2025). Empirical analysis of 47 transmission trees for 13 human pathogens found that self-organizing bursts of superspreading are common and that many trees lie near this critical boundary; over two thirds had posterior mean $k=\mu_\beta^2/\sigma_\beta^2$2 (Leonardi et al., 18 Sep 2025).

Correlation also alters depletion of susceptibles and the location of herd immunity. In the fully correlated gamma model of (Oz et al., 2020), where infectiousness and susceptibility are identified and both follow a Gamma distribution, the effective reproduction number declines as

$k=\mu_\beta^2/\sigma_\beta^2$3

and the herd-immunity fraction is

$k=\mu_\beta^2/\sigma_\beta^2$4

For $k=\mu_\beta^2/\sigma_\beta^2$5 and $k=\mu_\beta^2/\sigma_\beta^2$6, this gives $k=\mu_\beta^2/\sigma_\beta^2$7; for $k=\mu_\beta^2/\sigma_\beta^2$8, $k=\mu_\beta^2/\sigma_\beta^2$9 (Oz et al., 2020). The mechanism is selective depletion: the epidemic removes individuals who are simultaneously highly susceptible and highly infectious. Yet the same paper emphasizes a distinction between the herd-immunity point, where $k$0, and the final epidemic size; without interventions, substantial “after-burn” can occur after the threshold has been crossed (Oz et al., 2020). Correlation can therefore lower the herd-immunity threshold while leaving room for a large ultimate attack fraction.

6. Inference, control, and limitations

Superspreading complicates inference because overdispersion widens uncertainty in real time estimates of transmission. In the “disease momentum” framework, individual reproduction numbers are gamma distributed with dispersion $k$1, and daily cohort infectiousness is

$k$2

A transparent approximation links superspreading to the width of credible intervals for $k$3: $k$4 In Austria, prediction intervals from superspreading-aware models were about $k$5–$k$6 times wider than homogeneous alternatives, and $k$7 prediction-interval coverage increased from $k$8 in the Poisson-style model to $k$9 in the momentum model (Johnson et al., 2020). A plausible implication is that correlated superspreading would widen effective uncertainty still further, because dependence reduces the effective sample size beyond what is already implied by small $\mu_\beta = 0.18$0.

Control strategies exploit the fact that superspreading is concentrated in identifiable settings. A broad SARS-CoV-2 synthesis modeled “tail cutting” by preventing individuals who would otherwise cause more than 10 secondary infections from realizing that tail. With mean $\mu_\beta = 0.18$1, preventing half of those events reduced the mean to $\mu_\beta = 0.18$2, and preventing all of them reduced it to $\mu_\beta = 0.18$3 (Althouse et al., 2020). For known closed outbreaks, targeted prophylaxis can be stronger still. In the Diamond Princess case study, intranasal sprays combined with testing and social distancing had a $\mu_\beta = 0.18$4–$\mu_\beta = 0.18$5-day window of opportunity; with prophylactic efficacy and coverage greater than $\mu_\beta = 0.18$6, the average number of infections could be reduced by over $\mu_\beta = 0.18$7 (Booth et al., 12 May 2025). By contrast, at a short conference event with a heterogeneous face-to-face contact network, prophylactic use could roughly halve an individual’s probability of infection and reduce severe infections, but could not prevent all infections or onward community transmission (Booth et al., 12 May 2025).

Several limitations recur across the literature. County-level infectiousness inference often assumes homogeneous mixing within counties and independence of individual $\mu_\beta = 0.18$8 values, so estimated superspreading is a lower bound if detection is imperfect or cross-county mixing is appreciable (Pozderac et al., 2020). Enclosed-space aerosol models assume well-mixed rooms and fixed efficacy of filtration or prophylaxis over event durations (Kolinski et al., 2020). Network analyses of synergistic SIS, $\mu_\beta = 0.18$9-cores, or directed transmission typically omit fine-grained behavioral adaptation or use stylized degree distributions (Mizutaka et al., 2021). Correlated superspreading is therefore best understood not as a single settled model but as a general class of mechanisms through which transmission heterogeneity becomes organized, persistent, and epidemiologically amplified.