New work from the Turing-RSS Statistical Modelling and Machine Learning Laboratory, including Bayes4Health postdoc Brieuc Lehmann and co-PIs Chris Holmes and Sylvia Richardson, introduces a method to debias targeted testing data to obtain improved local-area estimates of COVID-19 prevalence in England.

Surveillance of the spread of SARS-CoV-2 (i.e. coronavirus) has largely been based on targeted tested schemes. For example in England, individuals have been asked to take a PCR test if they have a particular set of symptoms (high fever, continuous cough, loss of taste or smell). Similarly, healthcare professionals must get regularly tested in order to protect those that they're caring for. These tested groups are typically unrepresentative of the wider population of interest - they tend to have higher test positivity rates (number of positive tests / number of tests taken) compared to the true population prevalence. Someone who has COVID-19 symptoms is more likely to test positive than someone who doesn't have symptoms, and people working in care homes, hospitals, etc. are at a higher risk of catching coronavirus. This targeted testing data is routinely used to infer infection prevalence and the (now familiar) effective reproduction number, Rt, i.e. the average number of people an infected person goes onto infect.

This work develops a causal framework that provides debiased, local, weekly estimates of COVID-19 prevalence by combining targeted test counts with data from the Real-time Assessment of Community Transmission (REACT) Study. REACT is a randomised surveillance study that sends out 100,000 PCR tests to people across England every month. It provides an unbiased measure of prevalence, but the precision is low due to the relatively small number of tests (compared to the targeted testing schemes). The causal framework includes a bias parameter capturing the probability of an infected person being tested versus a non-infected person being tested, and transforms targeted test counts to debiased estimates of the true underlying local prevalence and effective reproduction number. These local estimates of Rt are indicative of one/two-week-ahead changes in the number of positive tests. The authors also that observed increases in estimated local prevalence and Rt reflected the spread of the Alpha and Delta variants in November 2020 and May 2021 respectively. The results illustrate how randomised surveys such as REACT can augment targeted testing to improve statistical accuracy in monitoring the spread of emerging and ongoing infectious disease. Full details can be found in the preprint (Local prevalence of transmissible SARS-CoV-2 infection: an integrative causal model for debiasing fine-scale targeted testing data), plus all the code needed to run the analyses is available at https://github.com/alan-turing-institute/jbc-turing-rss-testdebiasing.