António Câmara
I am a PhD candidate in the Department of Government at Harvard University. I study political representation and information environments across American political institutions, mass media, and social media. I develop and apply natural language processing and machine learning methods for social science measurement and causal inference. Current projects map and explain how politicians criticize and delegitimize powerful institutions, how conspiracy theories spread online and how their spread can be reduced, and whether and how local newspapers and online social movements carry ideological information to voters.
I am a National Science Foundation Graduate Research Fellow, a Theodore H. Ashford Fellow, and an affiliate of the Institute for Quantitative Social Science and the Center for American Political Studies. Before Harvard, I received a Bachelor of Science in Computer Science with a minor in Political Science from Columbia University.
I can be reached at acamara@g.harvard.edu.
Working Papers
Institutions substitute politics for violence, yet politicians have short-term incentives to attack the rules that constrain them. Scholars of democratic backsliding warn that these attacks are growing in number and worsening in kind; scholars of American populism contend that such rhetoric is a recurring style. Both camps are right, but about different kinds of attacks: criticism demands an institution use its power differently; delegitimization denies it should have that power. Across 15 million congressional floor speeches from 1873 to 2026, legislators today criticize institutions at the highest rate on record, 20 percent above the mid-twentieth-century peak. Delegitimization is rare and has fallen: 11 percent of attacks in 1873, 2 percent by 2017, 4 percent today. Legislators farthest from power supply the most attacks: the presidential out-party and ideological extremists. Yet nearly all criticize, and half delegitimize at least once. The change is gradual yet unexplained by the critical junctures or long-run processes that existing accounts invoke. I construct the measures with open-weight language models that extract and classify references to organizations, sectors, and rules across state, political, market, and epistemic domains, then bias-correct and validate them. In Congress, rising anti-establishment politics is loyal opposition.
We study statewide ballot measures across several states. We apply optimal classification, W-NOMINATE, and principal components analysis to a dataset of endorsements (n = 28,043) made by newspapers and interest groups, and the votes cast by citizens, to estimate the locations of newspapers, groups, and voters in a multidimensional policy space. We find that newspapers are much less “ideological” than previous research has assumed. We find evidence that, at least in California, newspapers have become more ideological over time, though not more extreme. Next, we study editorial board endorsement articles (n = 5,210) using two natural language processing methods: stance detection and document similarity. We find that editorial boards allocate minimal attention to providing opponents’ perspectives and roughly comparable attention to advocating for their preferred position and providing neutral context. Furthermore, we find that newspapers’ ideal points as estimated by optimal classification and W-NOMINATE do a poor job of predicting the one-sidedness of editorials. Finally, we find no evidence that California newspapers mirror expert groups’ public communications on domain-relevant propositions.
In text analysis, researchers often rely on random sampling techniques to select training data. However, random sampling can be risky, especially dealing with high-dimensional text data, as it tends to yield high variance in out-of-sample performance and may overlook important regions of the representation space. To address this challenge, we propose an alternative framework and algorithm that explicitly learns the underlying distribution of the text data and selects the most representative documents for model training. We demonstrate the effectiveness of our approach using both simulated datasets and four different real-world text corpora. We also discuss the scope of our method and its potential integration into more complex contexts, such as active learning and large language models.
Classic scholarship argues that local political behavior is guided by self-interest rather than partisanship or ideology, partly because local information environments often lack partisan and ideological information. When voters do have access to this information, do they sort along partisan and ideological lines, or do they continue to behave in self-interested ways? To answer this question, we turn to housing policy. We compile data on the partisan composition of state-level housing preemption policies and the ideological valence of YIMBY movement tweets between 2018–2024 to show that voters’ subnational information environments have recently become rich with partisan and ideological information about housing policy. We then use this to develop realistic information treatments in a survey experiment with residents of major metro-areas (n = 7,734). We find that this information increases partisan divisions on housing policy, and even persuades some homeowners to support housing reforms that negatively impact their self-interest.
Publications
As natural language processing systems become more widespread, it is necessary to address fairness issues in their implementation and deployment to ensure that their negative impacts on society are understood and minimized. However, there is limited work that studies fairness using a multilingual and intersectional framework or on downstream tasks. In this paper, we introduce four multilingual Equity Evaluation Corpora, supplementary test sets designed to measure social biases, and a novel statistical framework for studying unisectional and intersectional social biases in natural language processing. We use these tools to measure gender, racial, ethnic, and intersectional social biases across five models trained on emotion regression tasks in English, Spanish, and Arabic. We find that many systems demonstrate statistically significant unisectional and intersectional social biases.
Growing polarization of the news media has been blamed for fanning disagreement, controversy and even violence. Early identification of polarized topics is thus an urgent matter that can help mitigate conflict. However, accurate measurement of topic-wise polarization is still an open research challenge. To address this gap, we propose Partisanship-aware Contextualized Topic Embeddings (PaCTE), a method to automatically detect polarized topics from partisan news sources. Specifically, utilizing a language model that has been fine-tuned on recognizing partisanship of the news articles, we represent the ideology of a news corpus on a topic by corpus-contextualized topic embedding and measure the polarization using cosine distance. We apply our method to a dataset of news articles about the COVID-19 pandemic. Extensive experiments on different news sources and topics demonstrate the efficacy of our method to capture topical polarization, as indicated by its effectiveness of retrieving the most polarized topics.