A new preprint (Bayesian imputation of COVID-19 positive test counts for nowcasting under reporting lag) has just been arXiv’ed, involving Bayes4Health postdoc Brieuc Lehmann and PI Chris Holmes. This is the first piece of work to come out of the new (Turing-RSS Statistical Modelling and Machine Learning Laboratory) supporting the Joint Biosecurity Centre's COVID-19 response.

In the UK, it can take up to five days for ‘Pillar 2’ swab tests (i.e. tests for those displaying symptoms) to be processed and collated into a central system. This 'reporting lag' means we don't know exactly how many positive cases were picked up two days ago (for example), and often a simple moving average over the last 7 days is reported instead. The proportion of under-reporting for a given lag is somewhat predictable: based on historical data, it is possible to estimate what proportion of cases from two days ago have been reported today, and use this to estimate the total number of cases.

This work makes use of the stability of the under reporting process over time to motivate a statistical temporal model that infers the final total count given the partial count information as it arrives. By adopting a Bayesian approach that provides for subjective priors on parameters and a hierarchical structure for an underlying latent intensity process for the infection counts, one can obtain a smoothed time-series representation now-casting the expected number of daily counts of positive tests.

By explicitly modelling the reporting lag, the authors obtain more accurate estimates of the number of cases from a small number (n < 5) of days ago, compared to a baseline approach that uses a simple 7-day windowed average. This can provide authorities with a better idea of the progress of the COVID-19 epidemic, allowing for faster response to changes at the local authority level. Full details can be found in the preprint, plus all the code is available at https://github.com/alan-turing-institute/jbc-turing-rss-nowcasting.