RESUMO
While the fighting in the Syrian civil war has mostly stopped, an estimated 5.6 million Syrians remain living in neighboring countries. Of these, an estimated 1.5 million are sheltering in Lebanon. Ongoing efforts by organizations such as UNHCR to support the refugee population are often ineffective in reaching those most in need. According to UNHCR's 2019 Vulnerability Assessment of Syrian Refugees Report (VASyR), only 44% of the Syrian refugee families eligible for multipurpose cash assistance were provided with help, as the others were not captured in the data. In this project, we are investigating the use of non-traditional data, derived from Facebook advertising data, for population level vulnerability assessment. In a nutshell, Facebook provides advertisers with an estimate of how many of its users match certain targeting criteria, e.g., how many Facebook users currently living in Beirut are "living abroad," aged 18-34, speak Arabic, and primarily use an iOS device. We evaluate the use of such audience estimates to describe the spatial variation in the socioeconomic situation of Syrian refugees across Lebanon. Using data from VASyR as ground truth, we find that iOS device usage explains 90% of the out-of-sample variance in poverty across the Lebanese governorates. However, evaluating predictions at a smaller spatial resolution also indicate limits related to sparsity, as Facebook, for privacy reasons, does not provide audience estimates for fewer than 1,000 users. Furthermore, comparing the population distribution by age and gender of Facebook users with that of the Syrian refugees from VASyR suggests an under-representation of Syrian women on the social media platform. This work adds to growing body of literature demonstrating the value of anonymous and aggregate Facebook advertising data for analysing large-scale humanitarian crises and migration events.
RESUMO
Mobile sensing data has become a popular data source for geo-spatial analysis, however, mapping it accurately to other sources of information such as statistical data remains a challenge. Popular mapping approaches such as point allocation or voronoi tessellation provide only crude approximations of the mobile network coverage as they do not consider holes, overlaps and within-cell heterogeneity. More elaborate mapping schemes often require additional proprietary data operators are highly reluctant to share. In this paper, I use human settlement information extracted from publicly available satellite imagery in combination with stochastic radio propagation modelling techniques to account for that. I show in a simulation study and a real-world application on unemployment estimates in Senegal that better coverage approximations do not necessarily lead to better outcome predictions.