Algorithms for Political Methodology in the Information Age

Abstract

Contemporary politics is bound to an unprecedented volume of online information, and to the algorithms and social networks that deliver it. From millions of online campaign ads to billions of fake news social media posts, these novel phenomena are a vital frontier for political science, yet we require new methodological tools to understand them. Specifically, we require new methods that allow us to understand the content of these massive and exciting data sources, as well as methods that allow us to understand their fundamental strategic and behavioral dimensions. This dissertation comprises four papers that each offer a new method for these problems. The first paper introduces a new means to obtain solutions to LDA topic models that summarize the content of large text datasets in a manner that is provably better fitted to the text data, more interpretable, faster, and more amenable to causal inference than the current state of the art. This solves a set of open methodological problems that have been extremely well-studied since the popularization of topic models two decades ago. It also provides researchers with a first practical means to run topic models with provable guarantees and interpretability properties similar to those of linear regression and other workhorse methods in the political methodology toolkit. I demonstrate the advantages of this method by applying it to several well-studied datasets, as well as an original dataset of every text-based online political campaign ad on Google in the run-up to the 2020 US Presidential Election. The second and third papers describe the results of an academic research/data collaboration with Facebook, which yielded the first-ever academic access to Facebook's internal data on the billions of fake social media accounts that introduce and spread fake news across the social network. These papers show that despite the oft-changing content and effects of fake news, there exist fundamental and persistent strategic dynamics that characterize how campaigns of fake-news-sharing social media accounts strategically connect to a social network. In addition to their theoretical contributions, our models also contributed to the removal of billions of malicious fake Facebook accounts over the last year. The fact that these simple models consistently achieve state-of-the-art performance in an oft-changing setting dominated by far more complex machine learning algorithms suggests that the social interactions they capture are central to the strategic environment in which fake news propagates. The fourth dissertation paper leverages the content of online news to develop a novel measure of security dilemma dynamics. Specifically, we assemble a dataset of the full texts of virtually every online English-language article published about China, and we use plagiarism analysis to reconstruct the network of inter-media and government-media influences responsible for the spread of popular news `meme' characterizations of Chinese politics over time. We show how this approach constitutes a novel means to theorize the process by which security narratives emerge in international relations.