Measuring the Ideology of Audiences for Web Links and Domains Using Differentially Private Engagement Data

Abstract

This paper demonstrates the use of differentially private hyperlink-level engagement data for measuring ideologies of audiences for web domains, individual links, or aggregations thereof. We examine a simple metric for measuring this ideological position and assess the conditions under which the metric is robust to injected, privacy-preserving noise. This assessment provides insights into and constraints on the level of activity one should observe when applying this metric to privacy-protected data. Grounding this work is a massive dataset of social media engagement activity where privacy-preserving noise has been injected into the activity data, provided by Facebook and the Social Science One (SS1) consortium. Using this dataset, we validate our ideology measures by comparing to similar, published work on sharing-based, homophily- and content-oriented measures, where we show consistently high correlation (>0.87). We then apply this metric to individual links from several popular news domains and demonstrate how one can assess link-level distributions of ideological audiences. We further show this estimator is robust to selection of engagement types besides sharing, where domain-level audience-ideology assessments based on views and likes show no significant difference compared to sharing-based estimates. Estimates of partisanship, however, suggest the viewing audience is more moderate than the audiences who share and like these domains. Beyond providing thresholds on sufficient activity for measuring audience ideology and comparing three types of engagement, this analysis provides a blueprint for ensuring robustness of future work to differential privacy protections.