I am a Postdoctoral Researcher at ETH Zurich. I am mainly interested in political economy, media economics and organizational economics.
My research combines economics and machine learning to understand how information is produced, transmitted, and influences social and political outcomes.
I received my PhD in Economics from Columbia University in 2026.
Research
-
An Image is Worth 1000 Words? Partisanship and Attention in Videos
(Job Market Paper)
Abstract
Political news today is increasingly consumed in attention-scarce, video-based environments. Yet existing research on media partisanship remains text-centric and implicitly assumes long-form, attentive processing. This paper studies how partisanship operates in video news through the transmission of multimodal signals: images and text. I develop a multimodal measure of video partisanship to quantify partisan content in video news and decompose the contribution of each modality. I show that the informational strength of each modality depends on attention: images convey partisanship rapidly through affective cues, while text operates through substantive information that requires sustained exposure to accumulate. A survey experiment using real news footage shows how these properties shape viewers’ responses to political videos. Partisan images elicit immediate emotional responses even under brief exposure, whereas partisan text shifts policy attitudes only with sustained exposure. Overall, the results show how low-attention video environments reshape persuasion, by filtering out substantive textual information, and relying mostly on the emotional effect of visual signals. Efforts to improve media quality must account for the asymmetric roles of visual and textual content.
-
Mapping Economic Theory with AI
(with Elliott Ash & Sergio Galletta)
· Project website
Abstract
We develop a method for mapping and measuring how economic theory papers reach their conclusions. An agentic AI system represents each paper as a claim graph, which shows how model assumptions and intermediate claims support the reported results. In a random sample of 190 theory papers from four leading economics journals, removing the most consequential starting input leaves an average of 91 percent of weighted headline and major claims unsupported. Next, we benchmark error discovery against published corrections from 100 papers and use claim graphs to assess the consequences of repair. The graphs trace which conclusions depend on mistaken derivations and measure how much of their economic content is lost after correction. While errors in theory papers tend to affect support for a large share of a paper’s conclusions, they rarely change the main economic conclusions after the smallest valid repair.
-
The Mechanistic Fixed-Effects Topic Model: Recovering Corporate Culture from Employee Reviews
(with Dan Biderman, David Blei, Wei Cai, Amir Feder & Andrea Prat)
Abstract
Probabilistic topic models are widely used to uncover latent themes in observational text, but comparisons across groups are vulnerable to environmental variation in language. When documents are generated in different contexts, such as across industries or sentiment polarities, standard topic models conflate contextual differences with topic content, while bag-of-words representations discard contextual nuance. We introduce the mechanistic Fixed-Effects Topic Model (mFETM), an unsupervised generative model that explicitly separates invariant topic content from environment-specific deviations. To retain context and tone while preserving interpretability, mFETM defines topics as distributions over interpretable semantic concepts extracted from a large language model using sparse autoencoders. Applied to over 900,000 pro and con documents from Glassdoor reviews of S&P 1500 firms across 20 industries, mFETM improves held-out document perplexity while yielding topics that are substantially less confounded by review sentiment and industry vocabulary. In downstream firm-level analyses, mFETM recovers stable dimensions of corporate culture that reliably predict inclusion in Fortune’s 100 Best Companies to Work For. Finally, because the model operates in representation space, steering along estimated fixed-effect directions enables controlled counterfactual generation of workplace text across environments.
-
With or Without Conditions? Equity-Efficiency Tradeoffs in Municipal Transfers in Brazil
(with Luigi Caloi)
Abstract
To improve service delivery, central governments can tie intergovernmental transfers to local policy performance. While performance-based transfers incentivize local governments and generate efficiency gains, they also shift transfers from low- to high-capacity governments, creating welfare-reducing inequities. We study this equity-efficiency trade-off using a bundle of transfer reforms to Brazilian municipalities. When two states tied transfers to relative educational performance, student test scores rose substantially: moving from the 25th to the 75th percentile of per capita conditional transfers increased scores by 0.16 standard deviations. However, the reform also widened funding disparities across municipalities, which translated into disparities in municipal expenditures in multiple sectors. In contrast, contemporaneous reforms to unconditional transfers had negligible effects on student outcomes. We use a simple model of optimal transfers to interpret these findings. Our results suggest that the introduction of performance based transfers delivered large efficiency gains, limited welfare losses from inequities, and was welfare-enhancing. We find minimal evidence of multitasking distortions or score manipulation. Instead, we document improvements in the quality of education inputs and find suggestive evidence of reduced corruption.
Work in Progress
- Measuring Research Quality (with Elliott Ash, Sergio Galletta & David Zollikofer)
Resources
- Text and Images as Data in Social Science: slides · lecture video
Contact
Email: aciccarone@ethz.ch