Research
Peer-reviewed Publications
The Impact of Media Framing in Complex Information Environments
Political Communication (2025), 42(5), 757–773. Solo-authored.
2025 Kaid-Sanders Award: Best Political Communication Article
Abstract
To what extent do news frames influence public opinion? While a large body of experimental research suggests sizable effects, it is unclear how these findings translate to authentically complex information environments. I exploit a rare change in the immigration framing of the largest German tabloid, Bild, to estimate the precise impact of this shift on immigration attitudes using large-scale panel data. Despite a 42% increase in the emphasis on crime in Bild’s immigration coverage, I find a robust and precise null effect on immigration attitudes and several related variables. These findings highlight that framing effects materialize under specific scope conditions that need to be considered when generalizing from experimental results.Publication Replication Materials Award Interview
Attitude Manipulation and Voting Intentions
Political Behavior (2026), 48(1), 411–433. Solo-authored.
Abstract
Framing is a common strategy in political communication and a large body of research documents the substantial effects of this rhetorical device on political attitudes. However, opinion manipulation often represents merely an intermediate goal in a greater effort to affect citizens’ primary instrument of political influence: voting. From a rational model of voting, it follows that framing should have downstream consequences for voting intentions. In this manuscript, I outline a theory of when framing effects on policy attitudes should translate into voting intentions, drawing on rational choice as well as identity-based theories of voting behavior. I hypothesize that citizens should resolve cognitive dissonance emerging from an increased or decreased distance to a party’s issue position by either updating their inclination to vote for the party or by changing their issue position. Partisans should adapt their issue attitude in support of their party, whereas non-partisans should update their voting intentions to reflect the party’s issue distance.
I test these pre-registered hypotheses in a representative survey experiment fielded in Germany (N = 3004). The results suggest that while citizens’ attitudes are responsive to framing, this responsiveness does not translate into voting intentions. Citizens do adjust their attitudes to support their preferred party when learning about parties’ policy positions. However, attitudinal changes do not translate into voting intentions, irrespective of respondents’ attachment to a given party. These results underline the stability of voting intentions, even in the context of attitudinal change, and suggest that cue-taking is a more conscious process than is often theorized. Most importantly, my results contradict a mechanistic understanding of the connection between policy preferences and voting intentions.
Publication Pre-Registration Replication Materials
Work in Progress
Algorithmic Pre-Moderation Reduces Toxicity by Changing What Users Write: Evidence from a Randomized Experiment in Newspaper Comments
Under review. RCT on Der Standard evaluating the impact of algorithmic pre-moderation on online toxicity in newspaper comments. With Laura Bronner, Paco Tomas-Valiente, and Dominik Hangartner.
Presented at EPSA 2025
Abstract
Increasingly, online platforms have shifted from pre-publication content mo-deration—screening posts before they appear—to post-publication review that relies on user reports and subsequent moderator removal. Yet, rigorous evidence on how pre- and post-moderation affect discourse quality and participation remains limited. We report results from a randomized field experiment conducted with a leading Austrian newspaper featuring highly active user comment sections and operating under a default post-publication moderation regime. Articles were randomly assigned either to this post-moderation condition or a treatment that added algorithmic pre-moderation, screening comments before publication while leaving post-publication review unchanged. Across 410 articles and 115,657 submitted comments, assignment to pre‑moderation reduced the share of toxic comments among ultimately published content from 4.5% to 3.4%, a 23% decline (\(p\)-value = 0.005). Importantly, this reduction is driven primarily by a decline in toxicity among submitted comments, indicating that users change what they write, not merely which comments are removed before publication. Engagement—the number of comments and unique commenters per article—remained unchanged. Comparing the same users across articles shows that individuals submit less toxic comments under pre‑moderation than under post-moderation alone, consistent with users matching their contributions to the more civil tone of pre‑moderated discussions. As platforms increasingly rely on post‑moderation, our findings demonstrate the benefits of algorithmic pre‑moderation and provide guidance for platforms and policymakers seeking to improve public discourse by preventing toxic content from setting the tone.
Visible moderator engagement reduces online toxicity: Evidence from a randomized field experiment
Working paper. RCT evaluating the impact of visible moderator engagement on online toxicity on the largest Swiss news platform, 20Minuten. With Laura Bronner, Paco Tomas-Valiente, and Dominik Hangartner.
Presented at EPSS 2026
Abstract
Online moderation—whereby platforms delete or downrank harmful content—is generally invisible to most users, leaving online discussions feeling like spaces where no one is enforcing conversational norms of civility. This can lead users to infer that toxic comments are tolerated or even encouraged, making them more likely to post such comments themselves. To improve online discourse, we evaluate an intervention designed to increase the salience of moderation and make users perceive that conversational norms are enforced online: visible moderator engagement, whereby clearly identified moderators engage with ordinary users. To test its efficacy, we run a randomized field experiment together with the largest Swiss online newspaper, 20Minuten. We compare the comments submitted under articles which experience the usual invisible moderation (deleting harmful comments before publication), with those in articles in which moderators additionally engage visibly in the comments section, writing welcome posts, pinning good comments, and responding to both constructive and potentially harmful comments. We find that visible moderator engagement reduces both the number and the share of toxic comments submitted, without reducing engagement. We show that this reduction in toxicity is entirely due to behavioral change among users, rather than selection of less-toxic users into actively moderated articles. A complementary survey experiment with a separate sample suggests that visible moderator engagement works via two distinct causal mechanisms: (a) making users feel monitored, and (b) modeling and drawing attention to constructive online interactions. We find no evidence that users learn what content is rewarded, or that they fear punishments for transgressing. Together, these findings show that online discourse may be improved by making users feel that conversational norms of civility are being actively enforced.
Improving the Recognition of Democratic Norm Violations
Working paper. Survey experiment aiming to improve citizens’ recognition of democratic norm violations. Joint work with Markus Kollberg.
Presented at EPSA 2025, SVPW 2026, and the CIS Workshop Zurich 2026
Abstract
Why do citizens support illiberal political actors despite reporting a strong commitment to liberal-democratic ideals? We hypothesize that citizens might struggle to apply abstract liberal-democratic ideals when facing concrete choices simply due to a lack of understanding of how these ideals are challenged in the real world. We design an information intervention consisting of a video and an interactive quiz, informing respondents about liberal-democratic ideals and three different kinds of political transgressions on these norms.
We estimate precise and robust null effects of our information treatment on citizens’ ability to recognize democratic norm violations and their support for liberal and democratic values. Instead, we document in exploratory analyses that a substantial part of the electorate illiberal attacks on the judiciary. Furthermore, our evidence suggests that support for illiberal transgressions is driven by a perceived lack of liberal-democratic quality in the attacked institution. These findings challenge the contemporary understanding of citizens’ commitment to democracy and their support for illiberal actors. They have fundamental implications for researchers designing interventions to increase citizens’ commitment to democracy.
Debate Quality is an Emergent Property
Working paper. Introducing the hierarchical transformer for the measurement of deliberative quality in online discourse. Lead author. Together with Paco Tomas-Valiente, Lena Song, Dominik Stammbach, Laura Bronner, and Elliott Ash.
Presented at EPSA 2025 and CompText 2025
Abstract
We criticize current approaches to the operationalization and measurement of debate quality in political science and computer linguistics. Although debate quality is uniformly conceptualized as a property of the entire communicative exchange, it is almost invariably operationalized and measured at the level of individual contributions, such as speeches or online comments. We argue that debate quality emerges from relationships between different contributions to a debate and cannot be reduced to properties of distinct comments.
Drawing on data from three different online communities, we empirically show that operationalizing debate quality at the level of debates captures information that cannot be retained with the aggregation of comment-level properties. We then propose a Hierarchical Transformer architecture to capture the relationships between different debate contributions. We train such a model on synthetic annotations generated by GPT-4o and show that the hierarchical architecture outperform state-of-the-art long-input encoder models.
Measuring Rhetorical Similarity with Supervised Machine Learning
Working paper. Presenting a text-based measure of party accommodation for individual politicians, studying far-right accommodation in Austria, Germany, and the Netherlands. Solo-authored.
Abstract
I develop a fine-grained measure of party accommodation exploiting systematic measurement error in supervised machine learning algorithms. I validate the measure for the measurement of accommodation to radical right parties through established parties and politicians in Austria, Germany, and the Netherlands. Results indicate that the method produces valid estimates of parties’ and speakers’ rhetorical similarities to respective radical-right parties and outperforms existing similarity measures and scaling methods. I discuss possible further applications and limitations.
Frustrated Democrats and Support for Illiberal Actions
Currently fielding. Multi-country survey experiment studying how dissatisfaction with liberal-democratic institutions shapes democratic backsliding. With Markus Kollberg. Funded by a Diligentia Foundation research grant.
Presented at EPSS 2026 and the CIS WS 2026
Realignment on High-Stakes Issues: Evidence from Local Vaccination Behavior in Germany
Ongoing project.
Abstract
A large literature debates the extent to which citizens’ issue attitudes follow or lead parties’ policy positions. I argue that current research underemphasizes the impact of attitudes on voting intentions by mostly focusing on weakly held, variable issue attitudes. This issue is often exacerbated by the lack of attitudinal data preceding the politicization of political Issues. This study draws on data of local vaccination behavior from Germany to address this issue. I study whether parties’ positional shifts during Covid led individuals to change their vaccination behavior or adapt their party preferences. I show that changes in local party support are strongly correlated with historical vaccination rates, but only for the election following the onset of the Covid pandemic. In addition, I present evidence from an event study exploiting a local leaders’ public declaration not to get vaccinated. I find limited effects of this partisan cue on local Covid vaccinations. These results suggest that realignment is the norm rather than the exception on high-stakes issues.
Dissertation
Framing and Voting - The German Immigration Debate and the Effects of News Coverage on Political Preferences
Dissertation, Humboldt-Universität zu Berlin
Abstract
A large experimental literature on framing effects suggests that citizens form rather limited political preferences, open to severe manipulation. If citizens’ attitudes were always so easily malleable for media outlets and political actors, it would not constitute a very meaningful input for the democratic process. This dissertation asks how these experimental findings translate into complex, real-world news environments and whether news frames structure citizens’ voting intentions. It provides a clear conceptualization of frames, on which it builds a method to identify news frames automatically, and theorises a link between news frames and voting intentions. The dissertation presents original and secondary data, exploring the relationship of news framing, immigration attitudes, and voting intentions. Providing a broad overview of immigration framing in the German news media, it shows that neither immigration attention nor framing can explain the rise of the radical-right AfD. It then exploits a change in the immigration framing of Germany’s largest tabloid, Bild, showing that this shift had no effects on immigration attitudes or voting intentions among its readers. The final empirical chapter presents experimental evidence revealing that framing only affects voting intentions among rather uninformed citizens. The findings contribute to the study of framing and public opinion, suggesting that citizens’ attitudes are not as easily manipulated and the power of the news media more limited than often thought. Instead, framing effects take place under highly specific conditions, which are often not fulfilled. The emerging picture of public opinion is one of crystallized and resistant attitudes, which only respond to novel events. In other words: whoever gets to the voter first, wins. Politics, in this view, is a pattern of critical events following upon each other, each presenting a unique opportunity to change the dominant understanding of an issue.
Book Chapters
Die Nationalratswahl 2017 unter besonderer Berücksichtigung der Silberstein-Affäre
Chapter in: Weßels, B., & Schoen, H. (2021). Wahlen und Wähler. Analysen aus Anlass der Bundestagswahl 2017. Springer VS Wiesbaden. Co-authored with Julia Partheymüller and Jakob-Moritz Eberl
Public Outreach
The Public Discourse Indicator
A public dashboard tracking the quality of online discourse in Switzerland.
More
The public discourse indicator is a joint project of the Public Discourse Foundation and the Public Policy Group at ETH Zurich. It tracks the quality of online discourse in Switzerland over time by processing and classifying hundreds of thousands of comments each week. I lead the technical development of the indicator together with Laura Bronner and Killian Conyngham.
Measuring Media Bias with Cross-Domain Deep Learning
Work together with Tom Arend, looking into cross-domain learning to estimate newspapers’ leanings towards different political parties.
Abstract
The measurement of ideology is one of the major applications of text analysis in political science. However, researchers often face scarcity of available labelled data totrain supervised models for their specific domain. Manualannotation is costly and often severely affected by subjective bias. We propose the use of cross-domain learning to fine-tune transformer models on available, labelled political texts issued by political parties to obtain a classifierof political ideology. Using a unique dataset of newspaperarticles authored by politicians, we test such an application in the German context. Comparing transformer neural networks fine-tuned on different data sets of party pressreleases, we present evidence on the feasibility of such anapproach. This contributes to the broad literature on textanalysis in the political domain, informing researchers onthe limitations of training powerful deep learning modelson political language with scarce training data. The experiments conducted by the authors indicate that the trainingon party labels cannot be easily translated to the domainof newspaper articles but that supervised models are highlysensitive to minor changes in the inputs. It is thus recommended to refrain from using deep learning in the absenceof sufficient training data.
Silberstein und Kern - hat der Skandal der SPÖ geschadet?
Blog-post (Coauthored with Markus Wagner) on a scandal in the Austrian Lower House Election 2017; also covered by Austrian daily newspaper “Der Standard”.