Research

Working Papers

Social Influence in Online Reviews: Evidence from the Steam Store (Latest Version) (WP link)

• Winner of the Best Paper award of the Business and Industry Section of the Royal Statistical Society at NIE Conference 2024
• Presented at CODE @ MIT 2024, RES 2025

How does social influence affect consumer ratings? Using a dataset from the popular Steam gaming platform I investigate how quality judgements depend on pre-existing consumer assessments. In 2019, Steam introduced a new review system which decreased the exposure of users to previous ratings. Firstly, I find that user ratings are dependent on average ratings. The result is not due to selection, and is robust to a range of alternative specifications. Secondly, the effect is heavily asymmetric: individual reviewers are more negative when exposed to a lower average rating, but do not respond to a higher one. I rationalize these results with a model of reviewer behavior. Finally, using owner data, I estimate a structural model of game choice. A 1% increase in rating is equivalent to a 2 dollar price reduction. This suggests social influence has large implications for buyers and sellers.

Can Hate Speech Be Banned Online? The Effects of Shutting Down Toxic Forums on Reddit with Lily Shevchenko

• Presented at NICEP 2025, Warwick PhD Conference 2025, HEC PhD Conference 2025, KCL PPE Conference 2025, CODE @ MIT 2025, LSE-Warwick PolEcon PhD 2026

Is deplatforming effective in reducing toxicity on social media? To answer this question we study a policy change on Reddit in June 2020 which led to a simultaneous ban of thousands of forums containing hateful content, but not the users of these forums. We use data on the near universe of comments left on Reddit to examine the impact of the ban on user behaviour in a differences-in-differences design. We find that the most active users of banned subreddits comment more after the policy change and substitute to new forums in the weeks after the ban. The increase in activity persists in the long run, but is not associated with higher toxicity: instead, the comments left by affected users outside banned subreddits contain 20% fewer instances of hate speech. We do not find evidence that the policy leads to lower quality of engagement, negative spillover effects or recreation of banned subreddits elsewhere on the platform. Overall, the results suggest that moderation targeting toxic digital spaces can be effective in combating hate speech without lowering user engagement, and thus can be aligned with platforms’ incentives.

Work in Progress

Atomic Attention on YouTube: ``Brainrot” and Short-form Content

Does short form content really harm attention spans? Using the staggered rollout of YouTube Shorts across markets and channels, alongside rich video-level data including transcripts and attention heatmaps from 20 popular US media channels, I find that the introduction of short-form content significantly worsens user attention to their usual videos. Skipping at a given video slice increases by approximately 40% relative to the pre-Shorts mean, with across-video comparisons showing that longer videos are worst affected. Within-video analysis reveals that users disproportionately skip toward the beginning of videos, but not the end, suggesting this worseing attention has implications for information acquisition. The effect is sustained over time and robust to controlling flexibly for views and video content. These findings provide empirical evidence of attention externalities from short-form content exposure.