Work in Progress
Offline partisan segregation exceeds online echo chambers in the United States: Evidence from 1 million American voters
Revise & Resubmit, Nature Human Behaviour
- Presented at MPSA and APSA 2023
Abstract
Social media is often blamed for the creation of echo chambers. However, these claims fail to consider the prevalence of offline echo chambers resulting from high levels of geographic partisan segregation in the United States. Our article empirically assesses these online versus offline dynamics by linking a novel dataset of voters’ offline geographic partisan segregation extracted from publicly available voter files for 180 million US voters with their online network segregation on Twitter. We investigate offline and online partisan segregation using measures of geographical and online isolation of every matched voter-Twitter user to their co-partisans offline and online. Our results show that while social media users do tend to form politically homogeneous online networks, these levels of partisan sorting are significantly lower than those found in their geographical offline settings. Thus social media is providing users with networks that are more heterogeneous than their offline environments. We also find that Democrats are more ideologically isolated than Republicans in both settings. Our results contribute to the emerging literature on political communication and place the homophily of online networks in a broader context, providing novel evidence on partisan sorting both online and offline.Understanding Beliefs in Misinformation: Repetition, Partisan Signals and Bayesian Processing
Under review
- Presented at APSA 2025
Abstract
Partisan motivations and repeated exposure are prominent explanations for inaccurate beliefs, yet their interaction and relative importance remain poorly understood. We integrate these frameworks in a unified Bayesian model of belief formation and test its implications in four pre-registered experiments on political misinformation in the United States. Partisan motivated reasoning robustly shapes beliefs, persists over time, and operates for true, false, and entirely fabricated content. In contrast, evidence for repeated exposure—the “illusory truth effect”—is weaker: it is inconsistent across experiments, disappears after one day, and is stronger in non-political than political settings. We also evaluate interventions designed to mitigate these biases and find that warning labels on false information significantly reduce the effects of prior exposure. These findings clarify the relative roles of partisan motivation and repetition in belief formation and provide a structured framework for understanding and mitigating susceptibility to misinformation in politically charged information environments.The Partisan Effects of Information Control Policies: Evidence from the X Ban in Brazil
In Preparation
- Presented at APSA 2025
Abstract
What happens to information environments when partisan conflict over information control policies emerges in democracies? While a large literature examines information control under authoritarianism, democratic governments have increasingly intervened in major online platforms. We study a prominent case: Brazil’s 2024 national ban on the social media platform X. Using an event-study design, we estimate the causal effects of the ban and examine how partisan identity shaped responses. Drawing on a large sample of politically engaged users and ideal-point estimates of ideology, we find strong partisan asymmetries. Conservative users not aligned with the government were more likely to circumvent the ban, and right-leaning news domains became markedly more prevalent on the platform. We describe this dynamic as a “sorting ratchet”: the ban segmented the digital public sphere along partisan lines, with effects that persisted even after restrictions were lifted. When content moderation enforcement becomes a partisan issue, it may deepen polarization and reshape information environments in democratic settings.Memory and Ideological Bias in Conversations With Large Language Models: Experimental Evidence Across Four Countries
Data Collection
Abstract
We investigate whether interactions between users and Large Language Models (LLMs) produce ideological responses and the role of chatbots’ memory in their ideological leaning. We embed gpt-5.6-terra in an online survey experiment fielded to national samples in Argentina, Brazil, the United Kingdom, and the United States. Each respondent completes two sequential conversations, and we measure the ideological content of both the model and the user’s text at every turn. Between the conversations, we randomize whether the model propagates information from the first round into the second. Within conversations, we test whether the model’s expressed position converges toward the respondent’s, whether the range of ideological content it expresses narrows over time, and whether it exhibits a directional tilt independent of the respondent. Across conversations, we test whether a memory-primed model imports ideological content from an earlier charged exchange into a subsequent topic, including one that would ordinarily not elicit polarized responses.Criminal governance amid the COVID-19 pandemic in Mexico
Under review
Abstract
When do criminal groups govern through provision and coercion? During the COVID-19 pandemic, organized criminal groups (OCGs) in Mexico governed the communities under their influence through coercion and provision of goods. Existing research suggests these two tools are at odds. Drawing on an online survey with an embedded list experiment fielded to a national sample of Mexicans in late 2021, and a novel nine-level measure of local criminal competition, we test two preregistered hypotheses: that competition would raise coercive control and lower non-coercive control. We find that coercive social control rises with competition, but OCGs rely on non-coercive control actions even when competition is high. The two strategies also respond differently to the state: only non-coercive control varies with state capacity. These patterns reflect two distinct logics of social control: territorial contestation and legitimacy competition. We conclude that social control is best understood not as following mechanically from territorial position.Using Large Language Models and Text-Completion to Augment Web-Browsing Data
In preparation
- Presented at APSA 2025