A Model of Harmful yet Engaging Content on Social Media  

(with George Beknazar-Yuzbashev, Rafael Jiménez-Durán)

AEA Papers and Proceedings, 114 (2024): 678-83 [Published Article (gated)] [SSRN Article] [Citation & BibTeX]

Abstract: Why do social media users spend so much time consuming content that seemingly harms them? We build a simple model to argue that advertising-driven platforms can find it profitable to display content that harms users when it is complementary to their time spent on the platform. These incentives disappear, absent network effects, in the case of a subscription-based business model because harmful content reduces the willingness to pay for the platform. Our results warn against interpreting increases in engagement on social media as welfare increases.

Preempting Polarization: An Experiment on Opinion Formation 

(with Daniel Kashner)

Journal of Public Economics, 234 (2024): 105122 [Published Article (open access)] [Published Article PDF] [SSRN Article] [Citation & BibTeX]

Abstract: Blind adoption of opinions put forward by political parties and influential figures can sometimes be harmful. Focusing on cases where the partisan gap in policy support has not yet arisen, we investigate whether its formation can be prevented by encouraging prior active engagement with non-partisan information. To address this question, we recruited N=851 Republicans for a study about net neutrality, an issue largely unfamiliar to the electorate, which refers to equal treatment of all internet traffic. In a pre-registered experiment, we randomly changed the order in which the following two types of information were provided: (i) partisan, underscoring Republicans’ opposition and Democrats’ support, and (ii) non-partisan, where the participants evaluated factual arguments about the pros and cons of the policy. Despite holding total information constant, we find that those who saw the non-partisan block first donated 46% more to a charity advocating for net neutrality (p=0.001). The treatment effect persisted in an obfuscated follow-up study, conducted several weeks after the intervention. However, we do not find an effect on donations when repeating the main study with a sample of Democrats.

Do Social Media Ads Matter for Political Behavior? A Field Experiment

(with George Beknazar-Yuzbashev)

Journal of Public Economics, 214 (2022): 104735 [Published Article (gated)] [Citation & BibTeX]

Abstract: We exploit Facebook’s introduction of a filter hiding ads from the feed as a unique opportunity to study the effects of online ads on political behavior. In a pre-registered experiment, we randomly assigned participants to hide political ads (treatment) or alcohol ads (control) for several weeks preceding the 2020 US elections. We report an insignificant intent-to-treat effect of political ads on turnout (2.3 pp.), but we cannot rule out a sizable positive effect, with 95% confidence interval of [-2.8,7.4]. The result may mask important heterogeneity, with political ads making Democrats slightly more motivated to vote and Republicans – substantially less. We explore the reasons for this effect, such as natural variation in ad content: the majority of Facebook ads on users’ feeds skewed Democratic. Lastly, the effect on measures of affective polarization and informedness was negligible.

Working Papers

Toxic Content and User Engagement on Social Media: Evidence from a Field Experiment

(with George Beknazar-Yuzbashev, Rafael Jiménez-Durán, Jesse McCrosky)

Abstract: As much as forty percent of social media users have been harassed online, but there is scarce causal evidence of how toxic content impacts user engagement and whether it is contagious. In a pre-registered field experiment, we recruited participants to install a browser extension, and randomly assigned them to either a treatment group where the extension automatically hides toxic text content on Facebook, Twitter, and YouTube, or to a control group without hiding. As the first stage, 6.6% of the content displayed to users was classified as toxic by the extension relying on state-of-the-art toxicity detection tools, and duly hidden in the treatment group during a six-week long period. Lowering exposure to toxicity reduced content consumption on Facebook by 23% relative to the mean – beyond the mechanical effect of our intervention. We also report a 9.2% drop in ad consumption on Twitter (relative to the mean), where this metric is available. Additionally, the intervention reduced the average toxicity of content posted by users on Facebook and Twitter, evidence of toxicity being contagious. Taken together, our results suggest a trade-off faced by platforms: they can curb users’ toxicity at the expense of their content consumption.

To the Depths of the Sunk Cost: Experiments Revisiting the Elusive Effect

(with George Beknazar-Yuzbashev and Sota Ichiba)

Abstract: Despite being often discussed both in practice and academic circles, the sunk cost effect remains empirically elusive. Our model based on reference point dependence suggests that the traditional way of testing it by assigning discounts may not produce the desired effect. Instead, we evaluate it across the gain-loss divide by randomizing the price (low, medium, or high) of a ticket to enter a real-effort task and observing its effect on playtime. Despite a strong intervention we vary the sunk cost by $2 for a 14-minute task and the sample size of N=1,806, we find only a small effect (0.09 SD or 1.1 minutes). We further explore the economic applications of the sunk cost effect in a field experiment on YouTube in which we randomize whether the time until a pre-video ad becomes skippable is shortened (0 s), default (5 s), or extended (10 s). We report two results. First, the intervention had an insignificant effect on video engagement (the time spent following the ad segment). Second, we detect a sizeable negative effect on the extensive margin more users left before the video started in the extended treatment (5.3 pp. difference relative to the shortened treatment). Taking the results of both studies together, we offer a cautionary tale that applying even the most intuitive behavioral effects in policy settings might prove challenging.