Better the Devil You Know: An RCT on News Consumption (Job Market Paper)
Abstract: This paper investigates the causal link between the public’s self-selective exposure to like-minded partisan media and polarization. I first present a parsimonious model to formalize a traditionally neglected channel through which media selection leads to reduced polarization. In a world where the media heavily distorts signals with its own partisan preferences, familiarity with media biases is vitally important. By choosing like-minded partisan media, news consumers are exposed to familiar news sources. This may enable them to arrive at better estimates of the underlying truth, which can contribute to an alleviation of polarization. The predictions of this model are supported by experimental evidence collected from a South Korean mobile news application that I created and used to set up an RCT. The users of the app were given access to curated articles on key political issues and were regularly asked about their views on those issues. Some randomly selected users were allowed to select the news source from which to read an article; others were given randomly selected articles. The users who selected their news sources showed larger changes in their policy views and were less likely to have radical policy views—an alleviation of polarization—in comparison with those who read randomly provided articles. The belief updating and media selection patterns are consistent with the model’s predictions, suggesting that the mechanism explained in the model is plausible. The findings suggest that the designers of news curation algorithms and their regulators should consider the readers’ familiarity with news sources and its consequences on polarization.
Social Learning of Political Elites: A Natural Experiment in Iceland (Joint with Matt Lowe)
Abstract: Many legislative chambers are segregated along party lines, limiting cross-party interaction. Would there be less polarization if politicians were integrated? Or are the beliefs of political elites unchanging when exposed to alternative views? This paper tackles these questions by exploiting random seating in Iceland's national parliament. We introduce a novel measure of similarity of beliefs between any two politicians: cosine similarity of the language used in political speeches. Using this measure, we find an in-coalition effect: cosine similarity is greater for two MPs that share the same political coalition (government coalition or opposition) than for two MPs that don't. This suggests that the measure captures meaningful partisan differences in language. Next, we find that when two MPs randomly sit next to each other, their cosine similarity in the next parliamentary session (when no longer sitting together) is significantly higher, an effect that is roughly 16 to 25% of the size of the in-coalition effect. The fact that the effect persists after MPs are re-seated suggests that politicians are learning from their neighbors, not just facing transient social pressure. This learning however does not reflect the exchange of ideas “across the aisle” – the effects are large for neighbors in the same coalition group, at 29 to 53% of the in-coalition effect, with no evidence of learning from neighbors in the other group. Based on this evidence, integration of legislative chambers would likely slow down, but not prevent, the in-group homogenization of political language.
Research in Progress
Social Learning of Political Elites with Multidimensional Group Identity: Natural Experiments in the Parliaments of Norway and Sweden (Joint with Matt Lowe)
Abstract: In our previous paper on Icelandic Parliament (see above) we find strong in-group social learning between MPs who randomly sit next to each other and no evidence of out-group learning. In this paper, we test the generalizability of this result in other contexts. We exploit quasi-random seating in national parliaments of Norway and Sweden. In Norway and Sweden, MPs from the same region sit together, a different form of parliamentary integration than in Iceland. Within each division by region, the seats are determined by age and tenure (Sweden) or a score (Norway—the score is a function of vote share, calculated by the Sainte-Laguë Method). We employ regression discontinuity design (RDD) to identify social learning between MPs who sit next to each other, where the main outcome variable is language similarity between MPs. We also explore heterogeneity in social learning: Is out-group social learning facilitated in a context where MPs are often forced to collaborate for the good of the region? The results of this paper shed light on how the social integration of political elites affects polarization, especially with an additional dimension of group identity such as regional identity.
Donald Trump, Father of Five, Future Creator of the Wall: Candidate Information and Polarization of the US Voters (Joint with Cory Smith)
Abstract: In this paper we investigate whether voters react differently to different types of candidate information. Specifically, can partisan polarization of voters be alleviated by providing biographic information about candidates to which voters can relate, rather than information about the candidates’ policy stances? In a collaboration with BallotReady, a social venture for voter participation, we conduct an online field experiment in several US states during the 2018 general election. We randomize the type of information about candidates that we send to voters, along with reminder messages for impending election. From surveys, we collect information about the extent of split-ticket voting and subjective evaluation of candidates’ quality and favorability. We also observe whether and how much the voters seek to learn about other-party candidates. We track their clicks on URL links to webpages that contain candidate information and record the time they spend on those webpages.