Nowcasting COVID-19 statistics

In the midst of a pandemic, we want to make use of available information as soon as we get it. Because Sweden Covid-19 deaths are reported with up to 14 days delay, it is only about two weeks after a specific date do we know how many people actually died that day. Death statistics therefore always appear to be on a positive trend with very few death the last days. Similar illusions of always improving situations are visible in most countries’ reports of Covid deaths and case counts by their actual event date.

It is dangerous if policy makers or the public fall prey to this illusion, as it might motivate countries to open up too soon. A new pre-print paper proposes a statistical method to estimate the actual numbe of deaths based on the reported number of deaths. The proposed method has two important benefits: (1) It ensures everyone is aware of the reporting delay. (2) By including prediction intervals it allows readers to coordinate over the size of the uncertainty of the current state of the pandemic.

The proposed model is based on a statistical method from ecology called the "removal method". Compared to a prediction based on historical averages, the proposed model is a bit more accurate and precise. The reason the difference is not larger is that the decline in Swedish deaths has been slow, allowing a constant prediction to do well. But the proposed model is far more flexible to changing circumstances, and would allow for a sudden increase (or decrease) in deaths. The model can furthermore easily be adapted to the varying circumstances around each country's reporting delay.

Link to paper: https://arxiv.org/abs/2006.06840

Adam Altmejd, Stockholm School of Economics