What influences COVID-19 infection rates: A statistical approach to identify promising factors applied to infection data from Germany

Moritz Mercker; Uwe Betzin; Dennis Wilken

In order to evaluate possible factors locally driving infection rates, we applied the 'least absolute shrinkage and selection operator' (LASSO) method [14, 15] in combination with cross-validation to the data. LASSO is known to reliably perform model selection even if the total number of possible predictors is high [9, 16] . In particular, we applied LASSO separately to three different types of models (using β 1 , β 2 , respectively β 3 as the outcome variable, where for β 1 and β 2 we used the normal probability distribution. Since the used software (the R-package glmnet, c.f. below) is not able to perform Betaregression, we reformulated for the LASSO-step the value β 2 into a binomial variable instead ('smaller one' vs. 'one'), using a binomial error distribution. We tested the following variables: On the ADlevel we introduced Longitude, Latitude (as well as their interaction term), the date of the first reported COVID-19 infection (first infection), the local percentage of people older than 65 years (Age) (data from https://www-genesis.destatis.de), the population density (Density), and the local infection intensity (percentage of people infected by COVID-19 -infection density). On the coarser level of states, we introduced human mobility data. In particular, we used mobility data from Google collected by the type of activity, where activity (visits and length of stay) has been quantified at different types of places such as retail and recreation, groceries and pharmacies, parks, transit stations, workplaces, and residential -compared to a baseline level before the COVID-19 pandemic (variables have been named All rights reserved. No reuse allowed without permission.