Is the reproduction number currently 0.7, or 0.85, or 0.641? Was it bigger yesterday than today? A thread on real-time estimation and false precision... 1/

Most real-time estimation of R involves looking at the ratio of change from one generation of infection to the next (which takes about 5 days on average: https://www.eurosurveillance.org/content/10.2807/1560-7917.ES.2020.25.17.2000257). In reality, however, we rarely see the moment of infection, we just see cases appearing later... 2/

So the first task is to work out data to use. Should we estimate R from cases, or hospitalisations, or deaths? The answer, of course, is we should look at all these data sources, noting that each has different delays we need to account for 3/

Because most of these delays occur over a period of several days, and people are infectious for several days, it generates statistical uncertainty that makes precise daily estimates of R very difficult... 4/

One way to deal with the time issue is to use real-time social contact data (e.g. https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-020-01597-8) to measure how people are currently interacting, and hence estimate R. But we should still validate these estimates against other data sources such as case curves. 5/

There's also the issue of setting-specific transmission. Community transmission is currently low in many European countries, with outbreaks centred on healthcare and care homes. This means current case data doesn't mean what it meant two months ago. 6/

R is also an average value, so it doesn't make sense to focus on R when case counts get very low, because transmission is driven by super-spreading events, which become disproportionately important when infection numbers are small https://twitter.com/AdamJKucharski/status/1245336637497913345 7/

In summary: we generally have to estimate R from multiple datasets, each with its own notable uncertainties. So don't expect a highly precise estimate of R, or one that will change in an easily measurable way from day-to-day. 8/

There's more about R estimation methods on the @cmmid_lshtm epiforecasts page ( https://epiforecasts.io/covid/ ). 9/

Final technical note: if cases are under-reported, it doesn't affect R estimation as long as under-reporting is consistent (because R uses the ratio of change). However, detection is now improving in many countries, so need to be extra-cautious about interpreting R 10/10