“SHOW ME WHO YOUR FRIENDS ARE”: CONSTRUCTING A SOCIAL NETWORK USING MOBILE LOCATION DATA

Abstract

This thesis charts the path for the use of mobile location data in social network research. Location sharing by smartphone applications provides a significantly under-explored source of data for social scientists who have traditionally relied on survey responses to construct networks.

Objectives This thesis empirically explores the extent of political polarization. Specifically, it addresses three questions: (1) what do social networks look like for different political and demographic groups, (2) where do we observe interactions within and between groups, and (3) what are the factors the predict an interaction’s existence? On top of contributing to the academic literature on ideological segregation, I aim to create a foundation which future researchers can use – modifying and improving on my decisions along the way.

Data & Methods I use location markers from 180,000 smartphones in Polk County, Iowa. The locations are collected by a range of third-party applications and aggregated by a private company that shared them for research purposes. I use a smartphone’s night-time locations to infer its home address, and subsequently combine it with precinct-level results from the 2016 election, demographic traits from the American Community Survey estimates, and location characteristics obtained from Google Maps. After showing that the users present in the data are broadly representative of Polk County, I construct a data set of interactions with information about the users interacting and the location where they do so.

Results I develop a “proximity theory” supported by three strands of evidence. First, contact between users within the same group is significantly more likely than contact between users across groups. Second, users who reside in the denser parts of Polk County (i.e. in urban Des Moines) are over-represented in the interaction data set. I find that this happens because people in denser areas spend more time in proximity to others, which leads the algorithm to pick up “interactions” when they don’t exist. As such, groups living in denser areas exhibit interactions in places like grocery stores, whereas groups from sparser neighborhoods are overwhelmingly more likely to interact in social establishments – bars or restaurants. Third, distance is the strongest predictor of interactions – outweighing the estimated impact of alternative drivers such as shared group membership.

Conclusion This initial examination shows (1) evidence that proximity, rather than ideology, may be driving interactions, and (2) the potential of location data to contribute to social network research. The pursuit of similar research should keep in mind two considerations. First, this data offers a variety of privacy risks, and any future research should eliminate the potential of individual identification. Second, the results required a range of assumptions about the definition of group membership or interactions – they should be probed to ensure replicability.