This repositiory provides code and some explanation for the ILM-prop model that participates in the German COVID-19 NowcastHub.
To predict the number of hospitalisations we consider the reporting process of both reported COVID-19 cases and reported hospitalisations. Recall that the reporting date of a COVID-19 case is shared for both the case and its hospitalisation, i.e. the case and hospitalisation are linked through this date.
As hospitalisations are only available as
Decomposition of the daily reported hospitalisation incidences into the known incidences
$\color{#FF7900}H_{t,d}$ , i.e. the reporting triangle, and the future weekly increments$\color{#00747A} H_{t, d + 7 (k + 1)} - H_{t, d + 7 k}$ . The last increment might not be a weekly one, but we expect few cases to occur for such long delays.
More formally, denote by
where
To leverage known reported incidences, rewrite this as
where
Assuming that the proportions
and finally predict
In essence, this model is a regression of reported hospitalisations on reported cases.
As hospitalisation is affected by age, we perform this procedure for all available age groups separately and finally aggregate over all age groups to obtain a nowcast for all age groups combined.
This describes our point nowcast for
Denote by
Clone this repository and run
make dependencies # install necessary R packages
make data # prepare data from RKI case and hospitalisations
make submissions # re-create submissions
in a terminal. You'll need to have R installed to run this model.
Building all submissions may take a while, even on a decently powered machine.
The result will be stored in the data/processed
directory in the file submissions-ILM-prop.csv
which consists of all submissions stacked on top of one another. See the NowcastHub wiki for more information about the data format.
Note that there might be small deviations from the submitted nowcasts because every retrospective change of the data affects (due to the long horizon of 12 weeks) potentially a lot of submissions and we only provide the data at release of this repository rather than a full history of the data.