Nowcasting: Predicting Demographics and Crime with Facebook Ad Data

Abstract

As human interactions move increasingly online through mobile phones, internet and social media, they leave digital traces that provide a novel source of data for studying society and understanding patterns of human behavior. Such data is often available at finer spatio-temporal scales than traditional data from surveys and population censuses. This paper is a proof of concept study on the viability of using aggregate data from Facebook's advertising audience estimates for predictions of demographic composition and crime rates at the ZIP code level in the United States. The advantages of this approach include the large user base of this social media platform which enables it to capture information about sizable populations in real time. In particular, aggregate data on user interests by location are used to predict demographics such as age, education and income and annual crime rates for assault, burglary and robbery crimes at the ZIP code level. The analysis suggests that, in general, models using Facebook data with ensembles of trees perform better than linear models. Moreover, for predictions of crime rates, models using Facebook data achieve lower errors than models using demographic data with the highest gains made in predicting assault crime rates. However, when the Facebook data and demographic data were combined together in crime rates prediction models, the performance of the model in terms of prediction error improved further than models using either of these data sources alone. Potential limitations of this approach as well as implications for future work are discussed.