Estimating severity from multiple data sources using Bayesian evidence synthesis

As we prepare for a possible COVID-19 pandemic, understanding the severity of the
epidemic, i.e. the proportion of infections that result in a severe event such as
hospitalisation or death, is crucial to monitoring and predicting the burden of the epidemic
on healthcare services. This burden is measured by the number of people infected who
require primary care from general practitioners, hospital admission, respiratory support
and/or admission to intensive care. Severity is quantified by infection-severity and caseseverity
risks, namely the probability that an infection (whether with or without symptoms)
or a symptomatic infection (clinical case) leads to a severe event. These quantities are
challenging to observe directly from a single dataset, as it is not possible to detect and
follow-up every case in a population, particularly early in the epidemic when case counts
may miss many asymptomatic or mild cases. The problem is compounded by the fact that
for patients still ill in hospital, we have not yet had time to observe whether they will
recover or not. We propose combining information from multiple datasets, both on
individuals and aggregated counts, to estimate severity while accounting for the
challenges of missing cases and not yet observing outcomes.

This COVID-19 Rapid Response award is jointly funded (50:50) between the Medical Research Council and the National Institute for Health Research. The figure displayed is the total award amount of the two funders combined, with each partner contributing equally towards the project.

In preparing for a possible COVID-19 pandemic, estimates of severity, in particular of both
the numbers of infections occurring at different levels of severity and the infection- and
case-severity risks, are crucial to understand and predict the burden and impact of the
epidemic on healthcare services. Such estimates are most importantly needed by age and
risk group (e.g. defined by co-morbidities) strata, although in the early stages of an
epidemic, strata-specific information is rarely available. No single dataset can provide
enough information on its own to estimate severity, but estimation is feasible by
synthesising multiple datasets, such as: line-listing data from first few hundred type
studies; surveillance data including case counts, numbers accessing healthcare, and
numbers of deaths; cohort studies; and household studies. Such a synthesis needs to
account for biases inherent in each data source, including differential ascertainment by
severity level; and to account for the incomplete nature of the data, which, collected in real
time, are typically affected by censoring of final outcomes (recovery/hospital discharge or
death). We propose to make the best use of both individual- and aggregate-level data that
will become available, by using a combination of survival analysis techniques (e.g. curerate
mixture or competing risks models) and Bayesian evidence synthesis in a single
analysis to estimate severity in real time, as data accumulate over the course of the
epidemic, and once the epidemic is over.