NETWORK SCIENCE · MODELLING

Why is it difficult to control and predict SARS-CoV-2 spread?

Dr Akanda Ashraf · 2020 · all articles · back to profile

Epidemic curves look smooth in retrospect, which makes forecasting them look easier than it is. Transmission does not happen in a well-mixed population; it happens over a contact network whose structure is uneven, changing and only partially observable. That mismatch between the model's assumptions and the world's structure is where most forecasting error comes from.

Real contact networks are heavy-tailed. A small number of people have many more contacts than the average, so the same reproduction number can produce very different trajectories depending on who is infected early. Averages describe the population poorly when variance dominates, and interventions that look equivalent on paper — reducing contacts by a fixed fraction versus removing high-degree contacts — have very different effects.

Structure also creates delay and clustering. Communities are densely connected inside and sparsely connected between, so spread proceeds in bursts as the infection crosses from one cluster to the next. Combined with reporting lags and incubation, the signal a forecaster sees is a delayed, blurred version of the process being modelled.

Then there is feedback. Behaviour responds to the epidemic, and the epidemic responds to behaviour. Rising case counts change contact patterns, which change case counts. A model that treats behaviour as fixed is forecasting a system that no longer exists once its own output is published.

The practical implication is modesty about point forecasts and seriousness about structure. Network-aware models, explicit uncertainty and scenario ranges are more useful to a decision-maker than a single confident curve — and they make the assumptions being relied on visible, which is what allows them to be challenged.