Using ICU data to improve the real-time estimation of the effective reproductive number of the COVID-19 epidemic in 9 European countries

We replicate a recent study by the Imperial College COVID-19 Response Team (Flaxman et al, 2020) that estimates both the effective reproductive number, R _t , of the current COVID-19 epidemic in 11 European countries, and the impact of different nonpharmaceutical interventions that have been implemented to try to contain the epidemic, including case isolation, the closure of schools and universities, banning of mass gatherings and/or public events, and most recently, widescale social distancing including local and national lockdowns. The main indicator they use for measuring the evolution of the epidemic is the daily number of deaths by COVID-19 in each country, which is a better statistic than the number of identified cases because it doesn’t depend so much on the testing strategy that is in place in each country at each moment in time.

We improve on their estimation by using data from the number of patients in intensive care, which provides two advantages over the number of deaths: first, it can be used to construct a signal with less bias: as the healthcare system of a country reaches saturation, the mortality rate would be expected to increase, which would bias the estimates of R _t and of the impact of measures implemented to contain the epidemic; and second, it is a signal with less lag, as the time from onset of symptoms to ICU admission is shorter than the time from onset to death (on average, 7.5 days shorter). The intensive care signal we use is not just the number of people in ICU, as this would also be biased if the healthcare system has reached saturation (in this case, biased downwards, as admissions are no longer possible when all units are in use). Instead, we estimate the daily demand of intensive care, as the sum of two components: the part that is satisfied (new ICU admissions) and the part that is not (which results in excess mortality).

Thanks to the advantages of this ICU signal in terms of timeliness and bias, we find that most of the countries in the study have already reached R _t <1 with 95% confidence (Italy, Spain, Austria, Denmark, France, Norway and Switzerland, but not Belgium or Sweden), whereas the original methodology of Flaxman et al (2020), even with updated data, would only find R _t <1 with 95% confidence for Italy and Switzerland .