Are Large-Scale Data From Private Companies Reliable? An Analysis of Machine-Generated Business Location Data in a Popular Dataset

Large-scale data from private companies offer new opportunities to examine topics of scientific and social significance, such as racial inequality, partisan polarization, and activity-based segregation. However, because such data are often generated through automated processes, their accuracy and reliability for social science research remain unclear. The present study examines how quality issues in large-scale data from private companies can afflict the reporting of even ostensibly uncomplicated values. We assess the reliability with which an often-used device tracking data source, SafeGraph, sorted data it acquired on financial institutions into categories, such as banks and payday lenders, based on a standard classification system. We find major classification problems that vary by type of institution, and remarkably high rates of unidentified closures and duplicate records. We suggest that classification problems can affect research based on large-scale private data in four ways: detection, efficiency, validity, and bias. We discuss the implications of our findings, and list a set of problems researchers should consider when using large-scale data from companies.