In 2018, in the weeks before the US midterm elections, a team of economists led by Hunt Allcott of New York University and Matthew Gentzkow of Stanford University, along with Stanford collaborators Luca Braghieri and Sarah Eichmeyer, recruited 2,743 Facebook users in the United States for what would become one of the largest and most rigorous randomized controlled experiments ever conducted on the welfare effects of social media. The participants, all of whom had used Facebook for at least 15 minutes per day, were first asked how much money they would need to be paid to deactivate their Facebook accounts for a four-week period ending shortly after the midterm vote. The 61 percent of participants who would accept the deal for less than $102 were randomly divided into two groups. One was paid to deactivate. The other was paid an equivalent sum but allowed to continue using Facebook as before. The team then tracked both groups closely throughout the four weeks and after, using daily text messages, surveys, and public data on voting and Twitter use.
The results, published in the American Economic Review in March 2020 under the title “The Welfare Effects of Social Media,” produced findings substantially less dramatic than the most alarming popular discourse about social media but consistent with a meaningful negative effect on user well-being. According to the Stanford Institute for Economic Policy Research summary of the study, participants who deactivated their Facebook accounts reported significantly higher subjective well-being than the control group — happier, less anxious, less depressed, and more satisfied with their lives. They spent more time with family and friends offline. They were also significantly less politically polarized on issue-position measures, although their general dislike of the opposing political party was not measurably affected.
What “less polarized” actually meant
The reduction in political polarization is one of the more carefully-documented findings of the study, and one that requires some unpacking. According to the NBER working paper that preceded the published article, the research team distinguished between two kinds of polarization. The first is “issue polarization” — the extent to which people on opposite sides of the political spectrum hold different views on specific policy questions like immigration, taxes, healthcare, or gun control. The second is “affective polarization” — the extent to which people feel emotionally negative toward members of the opposing political party. Both forms of polarization are heightened by social media exposure to politically charged content, but the two phenomena are driven by somewhat different mechanisms.
The deactivation experiment found a meaningful reduction in issue polarization — roughly 0.16 standard deviations across the sample, which translates in practical terms to a small but real shift toward less extreme positions on policy questions. The effect on affective polarization, the emotional dislike of the other party, was smaller and less statistically reliable. The team interpreted this as evidence that Facebook’s effect on political views runs primarily through the content people encounter on the platform — political news, opinion pieces, posts shared by friends — rather than through any direct effect on cross-partisan animosity. Removing the content reduced the polarization. The affective component, more deeply embedded in long-term political identity, was less susceptible to a four-week behavioral intervention.
The well-being numbers
The well-being effects were similarly real but modest. According to an Inc.com summary of the study published shortly after the original NBER paper appeared in 2019, the deactivation group gained an average of about one hour per day of time previously spent on Facebook, with most of that extra hour spent on offline activities rather than on other social media platforms. They reported small but statistically meaningful improvements on standard measures of subjective well-being — about 0.09 standard deviations on the team’s composite well-being index, which combined questions about happiness, life satisfaction, anxiety, depression, loneliness, and similar measures. The team noted that the effect size, while clearly positive, was smaller than the popular discourse on social media harms might have suggested. Facebook was making participants somewhat less happy, but not catastrophically so. The platform’s negative impact on subjective well-being was real and measurable but, in absolute terms, modest.
The MIT economist Erik Brynjolfsson, director of the MIT Initiative on the Digital Economy, told the New York Times after the study’s release that the methodology was “the gold standard for how to do science.” The combination of large sample size, random assignment, multi-method measurement, pre-registration of hypotheses, and careful attention to placebo effects made the study substantially more rigorous than the observational studies that had previously dominated the social-media-and-wellbeing literature. The findings have been broadly replicated in subsequent work, including a much larger 2024 study by an overlapping research team that deactivated nearly 20,000 Facebook users and over 15,000 Instagram users for six weeks before the 2020 presidential election. The 2024 study, published in PNAS, found similar but slightly smaller effects on well-being and political knowledge.
Why people went back
The most counterintuitive finding of the Allcott-Gentzkow study was what happened after the experiment ended. Despite measurable improvements in well-being during the deactivation period, despite the increased offline time and reduced political polarization, the great majority of participants reactivated their Facebook accounts and resumed using the platform once the experimental payments stopped. The paper documented some persistent effects — treatment-group participants used Facebook less than they had before the study, with the effect persisting for at least nine weeks after deactivation ended — but the basic pattern was clear: people who had been paid to leave Facebook, had benefited measurably from the experience, and had been free to remain off the platform afterward overwhelmingly chose to return to it.
The reasons for the return are not obvious from the data alone, but the paper and subsequent commentary have suggested several plausible mechanisms. Facebook, even when it makes its users somewhat less happy on average, provides valuable functions that are difficult to replicate elsewhere. Users keep in touch with extended family and old friends through the platform. They get event invitations, group notifications, business contacts, news, and a steady stream of socially-relevant updates that would require substantial effort to obtain through other channels. The platform also exhibits classic habit-formation patterns — checking Facebook is reflexive for many users, and the habit re-asserts itself after a break. Additionally, the team measured that participants substantially revised downward their estimate of how valuable Facebook was to them after the deactivation experience. Their average willingness-to-accept for a future month-long deactivation dropped substantially. They returned, in other words, but with a more accurate sense of what they were returning to.
What the study did not show
The Allcott et al. findings have been widely interpreted as evidence that social media is harmful to users’ well-being, and the basic direction of that interpretation is supported by the data. The interpretation should be qualified, however, by what the study did not establish. It did not show that Facebook causes serious mental illness or addiction in typical users. The well-being effects, while real, were modest by clinical standards. It did not show that all social media is similarly harmful — Instagram, TikTok, Twitter/X, and other platforms have different content, different social dynamics, and different effects, and the 2018 Facebook study cannot be straightforwardly extended to those platforms. It did not test whether reducing Facebook use (rather than eliminating it entirely) would produce similar benefits with less disruption. And it did not address the more recent concerns about heavy social media use among adolescents, since the sample was limited to adults.
What the study did clearly establish is that Facebook, taken in aggregate across thousands of users, produces measurably negative effects on subjective well-being and political moderation, that those effects are large enough to detect in a rigorously-controlled experiment, and that users themselves often return to the platform anyway despite those effects. The combination — modest harm, demonstrated benefit from temporary disengagement, and persistent voluntary engagement — is roughly the same pattern that has been documented for many other kinds of mildly habit-forming behaviour, from snacking to television to alcohol. Facebook is not, in the Allcott-Gentzkow data, an exceptional case of digital catastrophe. It is, more modestly, a case of a popular consumer product that makes its users somewhat worse off than they would be without it, while remaining useful enough that they continue to choose it. Which is, perhaps, the most honest summary the data supports: not a moral panic, not a clean bill of health, but a small persistent cost that users are mostly willing to pay.
