Using location-based social media data to explain COVID-19 spread in Italy

Abstract

During crisis, social network data is valuable information for authorities and researchers. Since the COVID-19 outbreak, a significant amount of data, including mobility-related data, has been released from various agencies to support studies. This paper aims to check the suitability of mobility-related datasets in describing and predicting new cases, about the spatial dimension during the initial phase of the outbreak. We focus on rich anonymized datasets through Facebook – Data for Good program: colocation matrices, movement matrices and stay-at-home data. However, we also compare their usability with a traditional Origin-Destination matrix. Our test case is Italy, the second country hit after China, where the infection spread irregularly from a few northern provinces to rest of Italy and abroad. Our regression models explain the number of actual new cases at the provincial level (corresponding to NUTS-3) by three groups of variables: active cases proxying local infections, interprovincial mobility proxying the arrival of cases from outside, and the degree of people staying at home proxying infections from cohabitants. The variants among the models consist of different measures of interprovincial mobility, thus allowing us to confront them. The result is the inclusion of time-dependent mobility data improving the significance of model and is effective in explaining irregular rise of cases in different parts of the country. Moreover, colocation results as the best measure. From a policy perspective, results show that mobility restrictions help reduce the geographical spread of infection at the very beginning, but once the outbreak, the interprovincial mobility becomes less relevant.