Academic Research

Existing text scaling methods often require a large corpus, struggle with short texts, or require labeled data. We develop a text scaling method that leverages the pattern recognition capabilities of generative large language models (LLMs). Specifically, we propose concept-guided chain-of-thought (CGCoT), which uses prompts designed to summarize ideas and identify target parties in texts to generate concept-specific breakdowns, in many ways similar to guidance for human coder content analysis. CGCoT effectively shifts pairwise text comparisons from a reasoning problem to a pattern recognition problem. We then pairwise compare concept-specific breakdowns using an LLM. We use the results of these pairwise comparisons to estimate a scale using the Bradley-Terry model. We use this approach to scale affective speech on Twitter. Our measures correlate more strongly with human judgments than alternative approaches like Wordfish. Besides a small set of pilot data to develop the CGCoT prompts, our measures require no additional labeled data and produce binary predictions comparable to a RoBERTa-Large model fine-tuned on thousands of human-labeled tweets. We demonstrate how combining substantive knowledge with LLMs can create state-of-the-art measures of abstract concepts.

Existing approaches to estimating politicians' latent positions along specific dimensions often fail when relevant data is limited. We leverage the embedded knowledge in generative large language models (LLMs) to address this challenge and measure lawmakers' positions along specific political or policy dimensions. We prompt an instruction/dialogue-tuned LLM to pairwise compare lawmakers and then scale the resulting graph using the Bradley-Terry model. We estimate novel measures of U.S. senators' positions on liberal-conservative ideology, gun control, and abortion. Our liberal-conservative scale, used to validate LLM-driven scaling, strongly correlates with existing measures and offsets interpretive gaps, suggesting LLMs synthesize relevant data from internet and digitized media rather than memorizing existing measures. Our gun control and abortion measures -- the first of their kind -- differ from the liberal-conservative scale in face-valid ways and predict interest group ratings and legislator votes better than ideology alone. Our findings suggest LLMs hold promise for solving complex social science measurement problems.

Can a series of online conversations with a marginalized outgroup member improve majority group members’ attitudes about that outgroup? While the intergroup contact literature provides (mixed) insights about the effects of extended interactions between groups, less is known about how relatively short and casual interactions may play out in highly polarized settings. In an experiment in India, I bring together Hindus and Muslims for five days of conversations on WhatsApp, a popular messaging platform, to investigate the extent to which chatting with a Muslim about randomly assigned discussion prompts affects Hindus’ perceptions of Muslims and approval for mainstream religious nationalist statements. I find that intergroup conversations greatly reduce prejudice against Muslims and approval for religious nationalist statements at least two to three weeks post-conversation. Intergroup conversations about non-political issues are especially effective at reducing prejudice, while conversations about politics substantially decrease support for religious nationalism. I further show how political conversations and non-political conversations affect attitudes through distinct mechanisms.

For years WhatsApp has been the primary social media application in many countries of the Global South. Numerous journalistic and scholarly accounts suggest that the platform has become a fertile ground for spreading misinformation and partisan content, with some going so far as to assert that WhatsApp could seriously impact electoral outcomes, episodes of violence, and vaccine hesitancy around the world. However, no studies so far have been able to show causal links between WhatsApp usage and these alleged changes in citizens' attitudes and behaviors. To fill this gap, we conducted a field experiment that reduced users' WhatsApp activity during weeks ahead of the most recent Brazilian Presidential election. Our field experiment randomly assigns users to a multimedia deactivation, in which participants turn off their automatic download of any multimedia - image, video, or audio - on WhatsApp and are incentivized not to access any multimedia content during the weeks leading up to the election on October 2, 2022. We find that the deactivation significantly reduced subjects’ exposure to false rumors that circulated widely during the weeks before the election. However, consistent with the minimal-effects tradition, the direct consequences of reducing exposure to misinformation on WhatsApp in the weeks before the election are limited and do not lead to significant changes in belief accuracy and political polarization. Our study expands the growing literature on the causal effects of reducing social media usage on political attitudes by focusing on the role of exposure to misinformation in the Global South.

Social media is used by millions of Americans to acquire political news and information. Most of this research has focused on understanding the way social media consumption affects the political behavior and preferences of White Americans. Much less is known about Latinos’ political activity on social media, who are not only the largest racial/ethnic minority group in the U.S., but they also continue to exhibit diverse political preferences. Moreover, about 30% of Latinos rely primarily on Spanish-language news sources (Spanish-dominant Latinos) and another 30% are bilingual. Given that Spanish-language social media is not as heavily monitored for misinformation than its English-language counterparts (Valencia, 2021; Paul, 2021), Spanish-dominant Latinos who rely on social media for news may be more susceptible to political misinformation than those Latinos who are exposed to English-language social media. We address this contention by fielding an original study that sampled a large number of Latino and White respondents. Consistent with our expectations, Latinos who rely on Spanish-language social media are more likely to believe in election fraud than those who use both English and Spanish social media new sources. We also find that Latinos engage in more political activities on social media when compared to White Americans, particularly on their social media of choice, WhatsApp.

To what extent does the YouTube recommendation algorithm push users into echo chambers, ideologically biased content, or rabbit holes? Despite growing popular concern, recent work suggests that the recommendation algorithm is not pushing users into these echo chambers. However, existing research relies heavily on the use of anonymous data collection that does not account for the personalized nature of the recommendation algorithm. We asked a sample of real users to install a browser extension that downloaded the list of videos they were recommended. We instructed these users to start on an assigned video and then click through 20 sets of recommendations, capturing what they were being shown in real time as they used the platform logged into their real accounts. Using a novel method to estimate the ideology of a YouTube video, we demonstrate that the YouTube recommendation algorithm does, in fact, push real users into mild ideological echo chambers where, by the end of the data collection task, liberals and conservatives received different distributions of recommendations from each other, though this difference is small. While we find evidence that this difference increases the longer the user followed the recommendation algorithm, we do not find evidence that many go down `rabbit holes' that lead them to ideologically extreme content. Finally, we find that YouTube pushes all users, regardless of ideology, towards moderately conservative and an increasingly narrow range of ideological content the longer they follow YouTube's recommendations.

We test whether candidates move to the extremes before a primary but then return to the center for the general election to appeal to the different preferences of each electorate. Incumbents are now more vulnerable to primary challenges than ever as social media offers a viable pathway for fundraising and messaging to challengers, while homogeneity of districts has reduced general election competitiveness. To assess candidates' ideological trajectories, we estimate the revealed ideology of 2020 congressional candidates (incumbents, their primary challengers, and open seat candidates) before and after their primaries, using a homophily-based measure of domains shared on Twitter. This method provides temporally granular data to observe changes in communication within a single election campaign cycle. We find that incumbents did move towards extremes for their primaries and back towards the center for the general election, but only when threatened by a well-funded primary challenge, though non-incumbents did not.

Social networks play an important role in many political science studies. With the rise of social media, these networks have grown in both size and complexity. Analysis of these large networks requires generation of feature representations that can be used in machine learning models. One way to generate these feature representations is to use network embedding methods for learning low-dimensional feature representations of nodes and edges in a network. While there is some literature comparing the advantages and shortcomings of these models, to our knowledge, there has not been any analysis on the applicability of network embedding models to classification tasks in political science. In this paper, we compare the performance of five prominent network embedding methods on prediction of ideology of Twitter users and ideology of Internet domains. We find that LINE provides the best feature representation across all 4 datasets that we use, resulting in the highest performance accuracy. Finally, we provide the guidelines for researchers on the use of these models for their own research.

This article examines the news sharing behavior of politicians and ordinary users by mapping the ideological sharing space of political information on social media. As data, we use the near-universal currency of online political information exchange: URLs (i.e. web links). We introduce a methodological approach (and statistical software) that unifies the measurement of political ideology online, using social media sharing data to jointly estimate the ideology of: (1) politicians; (2) social media users, and (3) the news sources that they share online. Second, we validate the measure by comparing it to well-known measures of roll call voting behavior for members of congress. Third, we show empirically that legislators who represent less competitive districts are more likely to share politically polarizing news than legislators with similar voting records in more competitive districts. Finally, we demonstrate that it is nevertheless not politicians, but ordinary users who share the most ideologically extreme content and contribute most to the polarized online news-sharing ecosystem. Our approach opens up many avenues for research into the communication strategies of elites, citizens, and other actors who seek to influence political behavior and sway public opinion by sharing political information online.

Automated text classification has rapidly become an important tool for political analysis. Recent advancements in NLP enabled by advances in deep learning now achieve state of the art results in many standard tasks for the field. However, these methods require large amounts of both computing power and text data to learn the characteristics of the language, resources which are not always accessible to political scientists. One solution is a transfer learning approach, where knowledge learned in one area or source task is transferred to another area or a target task. A class of models that embody this approach are language models, which demonstrate extremely high levels of performance. We investigate the performance of these models in the political science by comparing multiple text classification methods. We find RoBERTa and XLNet, language models that rely on theTransformer, require fewer computing resources and less training data to perform on par with – or outperform – several political science text classification methods. Moreover, we find that the increase in accuracy is especially significant in the case of small labeled data, highlighting the potential for reducing the data-labeling cost of supervised methods for political scientists via the use of pretrained language models.

The role of the media in influencing people’s attitudes and opinions is difficult to demonstrate because media consumption by survey respondents is usually unobserved in datasets containing information on attitudes and vote choice. This paper leverages behavioral data combined with responses from a multi-wave panel to test whether Democrats who see more stories from liberal news sources on Twitter develop more liberal positions over time and, conversely, whether Republicans are more likely to revise their views in a conservative direction if they are exposed to more news on Twitter from conservative media sources. We find evidence that exposure to ideologically framed information and arguments changes voters’ own positions, but has a limited impact on perceptions of where the candidates stand on the issues.

The following report is intended to provide an overview of the current state of the literature on the relationship between social media; political polarization; and political “disinformation,” a term used to encompass a wide range of types of information about politics found online, including “fake news,” rumors, deliberately factually incorrect information, inadvertently factually incorrect information, politically slanted information, and “hyperpartisan” news. The review of the literature is provided in six separate sections, each of which can be read individually but that cumulatively are intended to provide an overview of what is known—and unknown—about the relationship between social media, political polarization, and disinformation. The report concludes by identifying key gaps in our understanding of these phenomena and the data that are needed to address them.