A Bayesian Approach For Estimating SARS-CoV-2 Setting-Specific Transmission Rates

In a recent damning report by the National Audit Office, it was revealed that the UK Government's Test and Trace programme aimed at combatting the Coronavirus pandemic, repeatedly failed to meet targets for contact tracing and test results, despite escalating costs of over £22 million. One such failing was the lack of data collection and analysis aimed at understanding setting-specific SARS-CoV-2 transmission, thus preventing the implementation of effective, data-driven policy. This has resulted in the grouping of disparate activities when determining lockdown rules, that do not share remotely similar transmission rates according to the few small-scale studies that exist.

In contrast, we demonstrate the use of an alternative methodology in which recipients of antigen tests complete a short survey detailing the activities they recently participated in. These are fed into a novel, first-principles-based Bayesian model, to infer where transmission is occurring. We benchmark the performance of the solution on simulated data in terms of convergence and runtime as the number of settings and observations increase, whilst evaluating the impact of model priors derived from expert opinion on our inference. Finally, we introduce a hierarchical variant of the standard model that allows us to incorporate interventions such as social distancing and mask-wearing into our model and intuit their effectiveness in each respective setting.

Our model is implemented in Python, comparing the use of two statistical computing frameworks (Stan and TensorFlow Probability) to determine the most computationally efficient approach. Reproducible examples of each implementation are made available through Google Colab.