Warped Words: How Online Speech Misrepresents Opinion

Abstract

I seek to resolve two seemingly contradictory facts of American politics: (1) most people hold moderate or mixed political views, and yet (2) online political discourse is (apparently) polarized. I investigate a theory that social media users falsify extreme views and/or self-censor moderate views, leading to a polarizing misrepresentation of opinion in online speech. The Introduction explores my motivating puzzle – the combination of attitudinal moderation and (perceived) online polarization – and connects it to relevant social science literatures. Chapter 1 observationally compares Twitter users' survey-reported political attitudes to their publicly-posted tweets, and the tweets of the people they followed (scaled using a text classifier I developed), finding evidence of preference falsification (though not necessarily in a polarizing direction), and of polarizing self-censorship by moderate users. Chapter 2 introduces a method – the “What Would You Say?” question and the Wordsticks model – to address limitations of the Chapter 1 analysis. This method combines conventional document scaling methods with survey-experimental causal inference. I validate and demonstrate the measure, and find evidence that preference falsification in offline conversations with extreme ideologues favors moderation, rather than polarization. Chapter 3 applies this method to estimate social media platforms’ effects on users’ speech, in a pre-registered experiment with a representative sample of Facebook and Twitter users. I find that most users self-censor political language online, and I find strong evidence that online discourse is polarized by moderates’ self-censorship (in the form of complete self-selection out of online political discourse). Planned analyses indicated no polarizing preference falsification, but exploratory analyses suggest Facebook and Twitter may differ in this regard. Chapter 4 introduces a research framework that can help identify how different platforms give rise to different kinds of discourse, using the open-source Mastodon software. I conduct a demonstration study, in which I successfully induce a microcosm of online political discourse using financial incentives. I outline a post-PhD research agenda using this framework. The Conclusion discusses the implications of my research. Self-censorship may dominate preference falsification because it requires less effort. Observers of social media should remember that posts do not accurately represent opinion.