Attenuated Renewal Models Improve Epidemic Forecasts in Sexual Networks

Forecasting the trajectory of an emerging infectious disease outbreak is one of the most demanding tasks in computational epidemiology, and it becomes even harder when transmission unfolds within a highly structured population. A new study published in PLOS Computational Biology by William D. Green, Christopher E. Overton, Thomas Ward, Ciara V. McCarthy and Sebastian Funk addresses exactly this challenge for epidemics spreading through sexual networks of men who have sex with men, a setting in which contact patterns are far from random. The researchers systematically compared five classes of short-term forecasting models, ranging from simple random walk approaches to semi-mechanistic attenuated renewal processes, and found that a specific modification of the classical renewal framework consistently delivered the most reliable predictions across both simulated and real-world epidemics.

The central difficulty the authors tackle is overdispersion. In many populations, the assumption that each infectious individual generates new cases at roughly similar rates breaks down dramatically. Within sexual networks in particular, a small fraction of individuals with very high numbers of partners can account for a disproportionate share of transmissions, while most individuals generate few or none. This heterogeneity, combined with assortative mixing, in which people tend to form partnerships with others who have similar levels of sexual activity, and with a mixture of one-time and recurrent partnerships, produces epidemic curves that violate the assumptions of standard forecasting tools. Models calibrated on early outbreak data can therefore be badly misled when the depletion of highly connected individuals changes the effective transmission dynamics over time.

Renewal models have long been a workhorse of epidemic forecasting. They describe the number of new cases at a given time as a random draw whose mean depends on the number of cases in previous generations, scaled by the reproduction number. This structure is flexible, requires relatively little data, and has performed well during influenza seasons and other outbreaks. But in an overdispersed network, the renewal equation’s implicit assumption of homogeneous mixing becomes a liability. Early in an epidemic spreading through a sexual network, cases concentrate among the most connected individuals, inflating the apparent growth rate. As those individuals become infected and, crucially, as the pool of highly connected susceptibles is depleted, transmission slows in ways that a standard renewal model cannot anticipate, leading forecasts to systematically overshoot.

Green and colleagues’ key innovation is to attenuate the renewal process with a reverse-sigmoid curve. The idea is elegant in its simplicity: the effective contribution of recent case counts to future incidence is scaled down over time according to a sigmoid-shaped function that declines as the epidemic progresses. This attenuation mimics the early depletion of highly connected nodes in the network. When the outbreak is young and cases cluster among individuals with many partners, the model retains more of the renewal signal; as the epidemic matures and those hubs are exhausted, the attenuation reduces the weight given to the observed case series, preventing the model from extrapolating an unsustainable growth rate into the future. The result is a semi-mechanistic model that captures a genuine network effect without requiring detailed knowledge of who is connected to whom.

To evaluate this approach rigorously, the team used stochastic epidemic simulations on synthetic MSM networks, which allowed them to generate many outbreak realizations under controlled conditions with known contact structures. They also analyzed empirical mpox case data from major cities, providing a real-world test case, since the 2022 mpox outbreak spread predominantly within sexual networks and exhibited exactly the kind of overdispersed transmission that motivates the study. Forecast performance was assessed using the continuous ranked probability score, or CRPS, a standard metric that evaluates the full predictive distribution rather than a single point estimate. The CRPS rewards forecasts that are both sharp, meaning concentrated around the eventual outcome, and well calibrated, meaning that uncertainty intervals contain the truth at their stated probabilities.

Across both the simulated epidemics and the empirical mpox data, the attenuated renewal models consistently outperformed the alternatives. The comparison set included random walk approaches, which essentially extrapolate recent trends with minimal structure, as well as standard renewal models without attenuation and other semi-mechanistic variants. In heterogeneous networks, the reverse-sigmoid attenuation produced forecasts that were more balanced and more reliable: they avoided the systematic overshooting that plagued unattenuated renewal models and maintained better calibration throughout the epidemic course. The improvement was not marginal. In a forecasting context where public health decisions hinge on whether an outbreak is accelerating, plateauing, or declining, the difference between a well-calibrated and a biased predictive distribution can translate directly into better-timed interventions.

What makes this result particularly valuable is its practicality. Detailed reconstruction of a sexual network, with information about individual partnerships, their timing, and their structure, is rarely feasible during an active outbreak, and even retrospective studies struggle to capture it comprehensively. The attenuated renewal framework sidesteps this data requirement. It needs only the case time series that health agencies already collect, yet it encodes a first-order correction for the network phenomena that distort naive forecasts. The authors note that distinguishing network effects from other factors that shape epidemic curves, such as behavior change, vaccination campaigns, or seasonality, remains challenging and usually requires additional data. The attenuation approach does not claim to disentangle these mechanisms; instead, it offers a robust forecasting heuristic that performs well even when the underlying causes of deceleration cannot be separately identified.

The mpox application underscores the timeliness of the work. The 2022 multicountry mpox outbreak caught many surveillance systems off guard precisely because its transmission dynamics differed from those of respiratory pathogens for which forecasting infrastructure had been built. Incidence rose rapidly within dense sexual networks and then declined in ways that were difficult to anticipate with models designed for homogeneous mixing populations. A forecasting tool that explicitly accounts for the depletion of highly connected individuals offers a template for future responses to pathogens that spread through similar contact structures, and the framework could plausibly be adapted to other overdispersed transmission settings beyond sexual networks, wherever superspreading-like dynamics drive early growth.

The study also contributes a methodological lesson for the forecasting community more broadly. Benchmarking multiple model classes on both synthetic and empirical data, and scoring them with a proper metric such as the CRPS, provides a clearer picture of when particular model structures help and when they hurt. The finding that a relatively simple structural modification, motivated by a mechanistic understanding of network epidemiology, can dominate both simpler and more elaborate alternatives suggests that the path to better outbreak forecasts often runs through targeted, theory-informed refinements rather than wholesale increases in model complexity. Random walk models, despite their simplicity, remain a useful baseline, and the value of the attenuated renewal approach is demonstrated precisely against such benchmarks.

As infectious disease threats continue to emerge in structured populations, tools like the attenuated renewal process fill an important gap between purely statistical extrapolation and full mechanistic network simulation. The work of Green, Overton, Ward, McCarthy and Funk shows that incorporating even a coarse representation of network depletion into a lightweight forecasting model yields more balanced and reliable short-term predictions. For public health agencies confronting the next outbreak in a sexual network or any other overdispersed transmission setting, the message is encouraging: robust forecasts may be achievable with the data already at hand, provided the models are built with the right structural assumptions about how connectivity shapes the epidemic’s course.

Subject of Research: Short-term forecasting of infectious disease epidemics spreading through overdispersed sexual networks using attenuated renewal process models

Article Title: Forecasting of infectious disease epidemics in sexual networks using attenuated renewal processes

Article References: Green, W. D., Overton, C. E., Ward, T., McCarthy, C. V., & Funk, S. (2026). Forecasting of infectious disease epidemics in sexual networks using attenuated renewal processes. PLOS Computational Biology, 22(10), e1013987. https://doi.org/10.1371/journal.pcbi.1013987

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1013987

Keywords: epidemic forecasting, renewal models, sexual networks, mpox, overdispersion, men who have sex with men, PLOS Computational Biology, CRPS, stochastic simulation, network epidemiology, public health, semi-mechanistic models