Chankyung Pak

Digital trace data have become one of the central pillars of media research methods. Despite the opportunities for better understanding individual users’ true behaviors in the personalized media environment, many scholars have pointed out the potential for bias in trace data collections, questioning the generalizability of findings based on them. In this study, we propose two statistical bias correction methods–Inverse Probability Weighting (IPW) and Type II Tobit, which are designed to remedy selection bias of inference from digital trace data donated by research participants. Applying these methods to Facebook take-out data, we demonstrate how the correction methods can change estimated effect sizes, which is important for the translation of academic findings into real-world impacts. We conduct two simulation studies, one under fully synthetic and another under partially simulated conditions, and find that Type II Tobit generally provides a more robust and cost-efficient correction method for digital trace data.

Political campaigns increasingly rely on Facebook for reaching their constituents, particularly through ad targeting. Facebook’s business model is premised on a promise to connect advertisers with the “right” users: those likely to click, download, engage, purchase. The company pursues this promise (in part) by algorithmically inferring users’ interests from their data and providing advertisers with a means of targeting users by their inferred interests. In this study, we explore for whom this interest classification system works in order to build on conversations in critical data studies about the ways such systems produce knowledge about the world rooted in power structures. We critically analyze the classification system from a variety of empirical vantage points—via user data; Facebook documentation, training, and patents; and Facebook’s tools for advertisers—and through theoretical concepts from a variety of domains. In this, we focus on the ways the classification system shapes possibilities for political representation and voice, particularly for people of color, women, and LGBTQ+ people. We argue that this “big data-driven” classification system should be read as political: it articulates a stance not only on what issues are or are not important in the U.S. public sphere, but also on who is considered a significant enough public to be adequately accounted for.

The visibility of news and politics in a Facebook newsfeed depends on the actions of a diverse set of actors: users, their friends, content publishers such as news organizations, advertisers, and algorithms. The focus of this paper is on untangling the role of this last actor from the others. We ask, how does Facebook algorithmically infer what users are interested in, and how do interest inferences shape news exposure? We weave together survey data and interest categorization data from participants’ Facebook accounts to audit the algorithmic interest classification system on Facebook. These data allow us to model the role of algorithmic inference in shaping content exposure. We show that algorithmic ‘sorting out’ of users has consequences for who is exposed to news and politics on Facebook. People who are algorithmically categorized as interested in news or politics are more likely to attract this kind of content into their feeds – above and beyond their self-reported interest in civic content.

To disseminate their stories efficiently via social media, news organizations make decisionsthat resemble traditional editorial decisions. However, the decisions for social media maydeviate from traditional ones because they are often made outside the newsroom and guidedby audience metrics. This study focuses on selective link sharing as quasi-gatekeeping onTwitter – conditioning a link sharing decision about news content. It illustrates how selectivelink sharing resembles and deviates from gatekeeping for the publication of news stories.Using a computational data collection method and a machine learning technique calledStructural Topic Model (STM), this study shows that selective link sharing generates adifferent topic distribution between news websites and Twitter and thus significantly revokesthe specialty of news organizations. This finding implies that emergent logic, which governsnews organizations’ decisions for social media can undermine the provision of diverse news,which relies on journalistic values and norms.