Research

We develop a deep learning framework for estimating ideal point models. Our estimation algorithm relies on a regularized autoencoder to approximate the posterior distribution of ideal points given the observed data and a researcher prior. The resulting approach is fast, scalable, and performs well in Monte Carlo simulations. It is particularly useful when combined with embeddings to estimate ideal points on massive, multimodal, unstructured datasets -- such as texts, images, audio, and video footage. We illustrate the model's versatility in two applications. First, we map the ideological positions of US Congress members using their voting records and speeches (1873-2016). Second, we analyze the ideological stance of political advertisers on Facebook and Instagram during the 2024 US elections based on the videos they pushed to users. A Python package DeepLatent is provided to support future applications.

Does the decline of local social infrastructure fuel far-right voting? I study closures of PMU cafés — traditional working-class gathering places in France — by linking administrative records on betting terminals to commune-level results from legislative and presidential elections (1997–2022). A regression discontinuity reveals no immediate electoral response, yet event-study estimates show effects growing to 1.28 percentage points after four elections. This temporal pattern points to gradual erosion rather than discrete shocks. Placebo tests using bakery closures, which lack sustained social interaction, show no causal effects, confirming that the mechanism is specific to social infrastructure. Effects are approximately three times larger in rural communes, where PMU cafés often constitute the only gathering place. These findings extend socio-cultural degradation theory to France and demonstrate that the political consequences of community decline accumulate slowly, distinguishing them from the immediate effects documented for visible economic shocks.

This article relies on a novel dataset of 686,439 speeches and written questions made at the European Parliament from 1999 to 2022 to shed light on the use of emotional rhetoric by politicians in the European Parliament. We use a Natural Language Processing method developed by Ash & Gennaro (2022) to scale the emotion of speeches and provide several stylized facts about the emotional load of debates at the European Parliament. In particular, we find that, controlling for context and country-specific variables, politicians from populist parties tend to be more emotional than politicians from mainstream parties.

Eventually, we provide two sets of evidence showing that populist politicians' emotional rhetoric is likely to be contagious. First, we find robust correlational evidence between mainstream politicians' level of emotion and the vote share obtained by populist parties from their country in the European election. Second, we find that mainstream politicians facing high competition from populist politicians at home seem to have increased the emotional load of their speeches more than those facing low competition at home. These two results call for additional work on the effect of populist politicians' entry into elected assemblies on mainstream politicians, as well as its potential effects on polarization and the quality of deliberation within elected assemblies.

I investigate the influence of social interactions among Members of the European Parliament (MEPs) on the similarity of language they speak on the floor. Using the quasi-random allocation of seats in the European Parliament, I find that sitting adjacently increases language similarity among MEPs by 6% within the same party and by 10% across different groups. Within-party peer effects are equally influenced by convergence in the topics discussed and in the phrases used to address them. In contrast, between-party convergence is driven solely by a more similar manner of addressing topics. Peer effects are markedly stronger among pairs of women, new members and pairs from the same member state. Using seating variation in the European Parliament’s venues (Brussels or Strasburg), I find persistent peer effects: MEPs who have previously sat together speak more similarly, even once they do not seat adjacently anymore. Lastly, I show that being surrounded by MEPs with diverse characteristics, such as gender, seniority in the EP or nationality, has varying effects on the degree of convergence towards their colleagues. In a survey of present MEPs, most respondents doubt the possibility of peer influence.

For much of the twentieth century, manufacturing sustained not only jobs, but also the collective life of the communities built around it. We study what happened when this economic and social infrastructure receded, combining harmonized census data with administrative records on income, elections, local amenities, and associations for more than 10,000 French municipalities between 1968 and 2016. We exploit differences in the timing and intensity of manufacturing decline across municipalities within the same department. Manufacturing decline led to persistent economic losses and demographic decline, including higher unemployment, lower incomes, slower population growth, aging, and out-migration. It also reshaped local political life: turnout declined, support shifted away from the left and, more recently, toward the far right. At the same time, local social infrastructure weakened. Association creation slowed, social isolation increased, and declining municipalities today host fewer amenities and meeting places. This social infrastructure mattered for political participation: where postwar church attendance was higher, turnout remained substantially more resilient despite similar economic losses.

The dynamics of Libor manipulation is studied through the lens of a Bayesian model in which an agent learns the intensity of the supervision it receives. The supervisor can be active or inactive, and the agent chooses a manipulation intensity each period that yields a short-term payment. While an inactive supervisor allows the agent to cheat indefinitely, an active supervisor can discover the fraud, all the more easily that the extent of the manipulation is important. If the fraud is discovered, the agent is sanctioned and the game ends. In such a framework, the more patient the agent is, the more the learning motive pushes them to manipulate to discover its environment. This experimentation value implies that a sequence of myopic agents observing the actions of their predecessors generates less manipulation than a single agent living an infinite number of periods. We characterise the agent’s optimal strategy, stopping probabilities and payoffs under a set of fairly general conditions, and discuss the Libor scandal in light of these results.