Research

Intersectionality is widely recognized as one of the largest contributions to the study of race, gender, and class across the academy. However, the quantitative operationalization of intersectionality within political science is often unsatisfactory. I provide evidence that the Bayesian Multilevel Model is an accessible and flexible tool for understanding intersectional dynamics in political behavior. Using both a synthetic simulation and a real-world example with the American National Election Survey (ANES), I show how Bayesian Multilevel Models increase our inferential understanding of group-based heterogeneity in public opinion and political behavior. Conventional techniques, such as interaction terms, frequently yield estimates that are obscured by considerable noise, making it challenging to discern meaningful patterns. In contrast, the Bayesian Multilevel Model excels at revealing underlying patterns in small sample-size environments.

This chapter reviews the major advances in accounting for intersectionality empirically and embracing methodological pluralism within Political Science and related Social Sciences. Intersectionality, or approaching identity categories rooted in structural power such as race, gender, and class as inseparable, remains a site of intellectual promise particularly because of its utility for explaining the big questions in American politics. I focus on intersectional quantitative methods as a site for new innovations as it is the natural step after demonstrating the current literature's advances of frameworks to operationalize intersectionality.

Twitter represented an invaluable space for sparking and mobilizing political movements. This study analyzes over 8 million tweets related to #MeToo over two years, aiming to provide new insights into the movement's dynamics and its relationship to fourth-wave feminism. Our findings challenge assumptions about consciousness-raising efforts, showing that politicized calls to action do not immediately materialize in online spaces. The study also highlights a lack of intersectional discourse in consciousness-raising discussions, emphasizing the need for broader considerations of gendered sexual violence.

Class is a known determinant of political attitudes and behaviors, yet it is often overlooked in quantitative intersectional research due to challenges in operationalization. This oversight stems from two main issues: inconsistent definitions of class in survey instruments and sparse data. In this paper, we propose defining class as a context-dependent latent variable, estimated through mixture models. Traditional methods typically isolate a single socioeconomic status (SES) or subjective social status (SSS) measure as an independent variable, but mixture models integrate multiple facets of SES and SSS, identifying the component of class most pertinent to the political outcome being studied. Coupled with intersectional approaches like Bayesian Multilevel Models, this framework allows for a more comprehensive representation of relevant identities in data sparse environments. We demonstrate our method with two empirical examples using 2020 American National Election Studies data, showing that the significance of SES or SSS elements varies depending on the outcome. Our results also indicate that not accounting for class in intersectional modeling leads to biased estimates.

Much of the existing COVID-19 research has focused on the politicization of the pandemic in the United States, where public opinion diverged along party lines in outcomes like public health guidelines, policy responses, and individual health behaviors. However, little attention has been paid to divergences in public opinion on perceptions of what groups were negatively impacted most by COVID-19.

Using original data linking social media digital trace data from Twitter and survey responses from 2020, we explore both traditional associations with COVID-19 polarization like Trump support and partisanship, and new factors such as traditional media and social media consumption. We ask how exposure to conservative media and social media influences individuals' beliefs about which racial and class groups were most negatively affected by COVID-19's economic fallout.

Combined survey responses and social media trace data is quickly becoming the gold standard in studying the influence of social media. This process of "data donation" asks survey respondents to provide researchers with their digital footprint on relevant platforms. Errors in inference can arise if people select into the sample based on characteristics related to the concept of interest being studied.

On surveys fielded through YouGov in 2022 and 2024, we asked respondents to donate Facebook data, YouTube histories, TikTok downloads, and install a custom web-browsing plug-in. We show that ideological conservatives under-donate, as do individuals who self-report exposure to sensitive or polarized content online. We demonstrate how these biases impact inference and offer a re-weighting solution using supervised machine learning approaches.

Pure independents, who represent about ten percent of Americans, are defined by the lack of partisan structure in their political attitudes, behaviors, and preferences. We demonstrate that many of these people are willing to reveal underlying partisan preferences through the partisan feeling thermometers. Accounting for these feelings reveals a clear and stable partisan structure to their attitudes. Leveraging existing cross-sectional, panel surveys, and original data, we demonstrate that: 1) Most pure independents have an affective lean. 2) Affective-leaning independents have distinctly partisan attitudes, behaviors, and preferences. 3) Affective lean is directionally stable over time. Building on these findings, we propose a new measure of party identification — Partisan Identity 9 (PID9). This measure builds on the traditional Party Identification 7 (PID7) measure by splitting the pure independent category into three categories. PID9 meaningfully improves model performance and provides powerful insights into the political attitudes, behaviors, and preferences of American independents.

How do antagonistic political actors strategically appropriate social justice narratives? Research on framing and countermovements has documented contestation over language, but has yet to systematically measure how and when narratives originating in one movement are taken up and repurposed by opposing groups on social media on a large scale. The first barrier to this is systematically identifying pro and counter movements in social media text data, and the second is identifying the narrative evolution itself. We address both with a tensor-based implementation of joint sentiment-topic modeling (TJST) that lets the researcher impose priors on both movement-associated language and the key accounts of each side, recovering movement positions and their narratives in a single estimation step. Because these priors are analyst-specified and inspectable, the approach is more traceable and label-efficient than black-box pipelines that chain a fine-tuned transformer or LLM stance classifier to a neural topic model.

Using Twitter data from the Dobbs v. Jackson decision, we first identify pro and anti position coalitions within the abortion rights discussion. We then identify core social justice narratives associated with the originating movement and measure when, how, and by whom these narratives are subsequently adopted by antagonistic actors. We benchmark TJST against a supervised pipeline that fine-tunes a BERT stance classifier and then applies BERTopic within each predicted side, showing that TJST recovers movement positions and narratives jointly, with interpretable priors and without needing labeled training data. By tracing narrative cooptation across groups over time, we capture moments in which the political right appropriates social justice language to reframe viral events and political claims.

How do social media environments influence public attitudes toward immigration? We link an eight-year panel survey (2016-2024) with Twitter (X) digital trace data of respondents to study how online political discourse shapes immigration attitudes and policy preferences. Our approach combines (1) cross-sectional survey analysis, (2) short-term attitude change models using panel data tied to social media exposure, and (3) models capturing cumulative effects over eight years of the panel. Exposure to immigration-related content is measured using BERT-based classifiers to identify topics of all tweets by accounts followed by respondents, including media, politicians, and non-elite accounts. We disaggregate this content based on ideological scores of content producers. In addition, we use hate speech detection models to identify hateful content, and keyword-based analysis to identify content with xenophobic terms. We operationalize two different outcomes, the ideological extremity of the respondent on immigration issues, and perspectives on undocumented immigrants. This design allows us to test whether exposure to right-leaning and/or xenophobic content over time is associated with respondents moving rightward on their stance toward immigration.