Estimating SARS-CoV-2 transmission in England from randomly sampled prevalence surveys

Abstract

Surveillance is a crucial component of a pandemic response, informing public health and wider policy. Unfortunately, it is often subject to reporting and behavioural biases, reducing its robustness. The CIS (Coronavirus Infection Survey) was a large-scale, longitudinal prevalence survey, with participants recruited randomly from the general population, thereby avoiding these issues. In this thesis, I consider how to best use the CIS data for surveillance. In particular, I consider the estimation of incidence (the number of new infections) and transmission (where infections are occurring) in England. This relies on knowledge of the duration of detectability (how long individuals test positive for). Therefore, I first consider how to estimate the distribution of this duration in the general population. I adapt a pre-existing model of how an individual's viral load changes over time for this purpose. I use this detailed model to exploit frequent (daily) sampling of individuals in a cohort of contact traced individuals. The cohort was small and the follow-up time short meaning that the tail of the duration distribution is poorly estimated. To overcome this issue, I combine the information with the larger and longer CIS dataset. The CIS data poses several challenges because it is doubly interval censored and some individuals in the cohort have infections that are never detected. I develop a survival analysis framework to analyse this data, including novel methodology to incorporate false negatives into the framework. This results in an estimate of the full duration distribution in the general population, not previously available. Finally, I use this estimate to infer the incidence and transmission of SARS-CoV-2 in England, using only the CIS data. I fully propagate the sampling error, accurately include the shape of the duration distribution, and estimate the quantities by age; previous work lacked at least one of these. Furthermore, I fit to only the CIS data, showing that such a survey is adequate to estimate these quantities.