Social media and mental health: the small number and what it hides

In 2019 two researchers at Oxford took three datasets covering 355,358 adolescents and ran the well-being question through every defensible version of the analysis they could construct, rather than the one version that would have supported a conclusion. Orben and Przybylski's specification curve found a negative association between digital technology use and adolescent well-being. At the outside it accounted for 0.4% of the variance.

That is the number most arguments about this subject are conducted without.

A year later Odgers and Jensen surveyed the field for an annual research review and got to somewhere similar by a different road. Across the meta-analyses and the large preregistered cohorts, the associations are small, they cannot separate cause from effect, and they are unlikely to matter clinically.

So the headline is dull. What makes it worth going past is that an average taken over a population in which almost everybody uses these products says very little about any particular person inside it. A negligible mean and a badly damaged minority produce the same figure.

What happens when people stop

Correlation cannot fix direction. Miserable people might simply scroll more. Three studies get at that differently, and they agree.

Allcott, Braghieri, Eichmeyer and Gentzkow recruited 2,743 people and asked each what they would need to be paid to give up a large social network for four weeks. The 61% who would do it for under about a hundred dollars were then randomly assigned either to deactivate for the four weeks before the 2018 US midterm elections or to carry on as normal, which leaves an impact sample closer to 1,600 than to the recruited figure. Deactivation raised subjective well-being. It also cut factual news knowledge, lowered political polarisation, moved time offline, and left people using the product less after the experiment had ended and nobody was paying them anything.

Hunt, Marx, Lipson and Young ran something smaller and stricter. They took 143 undergraduates and randomly assigned them either to hold their use to about ten minutes per platform per day for three weeks, or to carry on as usual. Loneliness and depression fell in the capped group relative to controls. Anxiety and fear of missing out fell in both groups, which the authors put down to the self-monitoring the study imposed on everybody. Watching your own use did some of the work that cutting it did.

The third study did not intervene at all. Braghieri, Levy and Makarin noticed that a social network had reached US college campuses one at a time, which is about as close to a randomised rollout as the world normally offers, and compared student mental health on each campus before and after it arrived. Mental health got worse. Self-reported academic impairment went up. When they went looking for the mechanism they found social comparison, not hours.

Every one of those has a hole in it. Nobody can blind a deactivation study, so participants always know what is being measured, and the follow-up periods are weeks when the question is about years. Three different holes, one direction.

Pooling the experiments confirms that direction and deflates it at the same time. Burnell and colleagues meta-analysed 32 randomised restriction trials covering 5,544 people and found that cutting back does improve subjective well-being, at g = 0.17: real, consistently signed, and small. Two things inside that number matter more than the number. Every sample was college students or adults, average age 23, so none of it is evidence about adolescents. And nothing moderated the effect — not how long the restriction ran, not whether people abstained outright or merely cut down — which is hard to reconcile with any story where the harm scales with hours. Their own conclusion is that restriction is probably not the most effective way to improve well-being.

Screen time is a blunt variable

Ask how long somebody spent and you get a weak signal. Weak, but not nothing — the three-hour association below is a time measure, and small correlations between hours and low mood turn up repeatedly. Ask what somebody was doing, when, and with whom, and the answers sharpen. Maheux and colleagues put the structural version of this in a 2025 review: social media is not one exposure, most of the literature measures time because time is the easy thing to measure, and the variation worth having is in specific components and in who is using them.

Comparison

Verduyn, Ybarra, Résibois, Jonides and Kross drew the line the field has largely organised itself around since. Passive use, meaning scrolling and reading and looking, goes with lower well-being. Active use, meaning talking to people you actually know, goes with higher. Their proposed mechanism is social comparison and the envy that follows it, which is where the campus rollout study independently ended up from a completely different dataset.

The distinction has not held up as cleanly as the field's enthusiasm for it. Godard and Holtzman meta-analysed 141 studies covering about 145,000 people and found most active-and-passive associations negligible, under r = .10. The exceptions do not line up tidily either: active use went with more online support and higher well-being, and also with more anxiety; passive use went with more perceived online support rather than less. Passive use looked worse in general feeds but not inside groups, and age mattered. The mechanism the 2017 review named is still standing. The rule people extracted from it — passive bad, active good — is not.

Fardouly and Vartanian's review of the body image research is unusually candid about where its own evidence runs out. The correlational and longitudinal work linking use to appearance concerns in young women and men is consistent, and appearance comparison looks like what carries it. A brief experimental exposure in a laboratory did not reproduce the effect. Whatever is going on happens over months, and an hour in a lab cannot manufacture it.

Sleep

Alonzo, Hussain, Stranges and Anderson reviewed 42 studies of 16 to 25 year olds and concluded that sleep is a partial mediator: heavy use degrades sleep quality, and that degraded sleep carries part of the association with depression and anxiety. It is among the least contested findings here and one of the more useful, because it points at when people use these things rather than how much. The review's own caution belongs with it: 36 of its studies were cross-sectional and only six prospective, and the authors say directionality and strength both still need establishing.

Reward

Lindström and colleagues fitted standard reinforcement-learning models to more than a million posts from over 4,000 people. Posting fits reward learning quantitatively as well as qualitatively: people post at the rate that maximises the average rate of social approval coming back, which is how a foraging animal behaves. That is a smaller claim than "addictive" and a far better supported one. An unpredictable schedule of social approval is a reinforcement schedule and does what reinforcement schedules do.

Harassment

van Geel, Vedder and Tanilon pooled 34 studies covering 284,375 young people on suicidal ideation, and nine covering 70,102 on suicide attempts. Peer victimisation roughly doubles both — an odds ratio of 2.23 for ideation and 2.55 for attempts — and cyberbullying was more strongly related to ideation than traditional bullying was.

Those are the largest associations anywhere in this piece, by a distance. They are also pooled from observational studies, of peer victimisation in general rather than of social media in particular. So the size is well established and the causal reading is still inference. It is a more comfortable inference than most here, because the alternative direction is harder to tell, but it is inference.

Who it actually happens to

The World Health Organization's European office surveyed almost 280,000 adolescents across 44 countries in 2022 and found 11% showing signs of problematic use: unable to control it, letting other things go, taking the consequences at home and at school. Four years earlier the same survey put it at 7%. Girls were at 13%, boys at 9%.

Read what that measures, though, because it is easy to inflate. It is a pattern of behaviour, not a diagnosis and not an attribution: a single cross-section, in which depression, anxiety and social difficulty can drive problematic use at least as readily as they follow from it. The 11% is not a count of adolescents harmed by social media, and anyone using it that way, in either direction, is borrowing authority the survey did not issue.

The same report found something that gets quoted far less often. Heavy users who were not problematic users reported stronger peer support than light users did. Whatever separates the 11% from everybody else, it is not the number of hours.

The US Surgeon General's 2023 advisory is a public-health document rather than a study, and it collects the observational evidence into a blunter shape: adolescents spending more than three hours a day on these products carry roughly double the risk of poor mental health outcomes. The same advisory is explicit that benefits exist, that the evidence has real gaps, and that children should not be the ones required to prove the products are safe.

Those benefits are documented rather than conceded for balance. Naslund, Bondre, Torous and Aschbrenner reviewed what these networks do for people who are isolated, stigmatised, or living with mental illness: peer support and access to a community that frequently exists nowhere else in their lives. An account of the harm with no room in it for that is not finished.

What nobody has shown

Vuorre and Przybylski went looking for the population-level signal. Across countries, over the two decades in which the internet and then mobile broadband arrived, they found small and inconsistent movement in well-being and mental health indicators, and nothing resembling the trace a global harm ought to leave. The paper is contested. A published critique sets out one conceptual and three methodological objections to it, and that argument is still running.

Three further limits apply to nearly everything above.

Self-reported use is unreliable, and it is worse than that sounds. Parry and colleagues meta-analysed 106 effect sizes and found that what people say about their own media use correlates only moderately with what the logs record, and that measures of problematic use, the ones carrying most of the alarm, match the logs worse than ordinary use does. A good deal of the apparent disagreement in this field is measurement error.

Direction stays hard. Almost every design capable of establishing causation buys that capability with a narrow window, an unusual population, or an intervention nobody would ever encounter outside a study.

And two substitutions do most of the work in the public argument. "No consistent global effect" gets read as "no harm". "Harm to some" gets read as "harm to all". Neither follows.

Building one of these

The strong version of the alarm is not supported. A shorter list is, and every item on it is a decision somebody makes while designing rather than a warning to bolt on afterwards.

Comparison has the best evidence of any mechanism here, arriving independently from a review and from a natural experiment. That makes public counts, follower totals and anything else that ranks people against each other a design decision with a literature attached, rather than a default to inherit. Sleep has the clearest pathway, and it is about when a product is used rather than how much. An unpredictable schedule of social approval behaves like the reinforcement schedule it is, and declining to build one is available.

Passive consumption is the weakest of those and I would not build a roadmap on it. The direction survives, the meta-analysis says the effect is mostly negligible, and "more conversation, less scrolling" is a preference with some evidence near it rather than a finding.

Heterogeneity is the hardest part, and it is where I can tell you least. A small average is consistent with a minority being badly hurt. It is equally consistent with nobody being hurt very much. Nothing above settles which, because the studies that could — following particular people, with particular vulnerabilities, through particular features — are the ones the field is only now starting to run. What follows for design is not a prescription but a caution about method: the median user is the one measurement finds easily, and is not the one the safeguards are for.

None of this is medical advice, and none of it replaces talking to somebody. If your own use is making your life worse, that is worth raising with a doctor or a local mental health service.

Sources

1. Orben, A., & Przybylski, A. K. (2019). The association between adolescent well-being and digital technology use. Nature Human Behaviour, 3, 173–182. doi.org/10.1038/s41562-018-0506-1
2. Odgers, C. L., & Jensen, M. R. (2020). Annual Research Review: Adolescent mental health in the digital age — facts, fears, and future directions. Journal of Child Psychology and Psychiatry, 61(3), 336–348. doi.org/10.1111/jcpp.13190
3. Allcott, H., Braghieri, L., Eichmeyer, S., & Gentzkow, M. (2020). The Welfare Effects of Social Media. American Economic Review, 110(3), 629–676. doi.org/10.1257/aer.20190658
4. Hunt, M. G., Marx, R., Lipson, C., & Young, J. (2018). No More FOMO: Limiting Social Media Decreases Loneliness and Depression. Journal of Social and Clinical Psychology, 37(10), 751–768. doi.org/10.1521/jscp.2018.37.10.751
5. Braghieri, L., Levy, R., & Makarin, A. (2022). Social Media and Mental Health. American Economic Review, 112(11), 3660–3693. doi.org/10.1257/aer.20211218
6. Verduyn, P., Ybarra, O., Résibois, M., Jonides, J., & Kross, E. (2017). Do Social Network Sites Enhance or Undermine Subjective Well-Being? A Critical Review. Social Issues and Policy Review, 11(1), 274–302. doi.org/10.1111/sipr.12033
7. Fardouly, J., & Vartanian, L. R. (2016). Social Media and Body Image Concerns: Current Research and Future Directions. Current Opinion in Psychology, 9, 1–5. doi.org/10.1016/j.copsyc.2015.09.005
8. Alonzo, R., Hussain, J., Stranges, S., & Anderson, K. K. (2021). Interplay between social media use, sleep quality, and mental health in youth: A systematic review. Sleep Medicine Reviews, 56, 101414. doi.org/10.1016/j.smrv.2020.101414
9. Lindström, B., Bellander, M., Schultner, D. T., Chang, A., Tobler, P. N., & Amodio, D. M. (2021). A computational reward learning account of social media engagement. Nature Communications, 12, 1311. doi.org/10.1038/s41467-020-19607-x
10. van Geel, M., Vedder, P., & Tanilon, J. (2014). Relationship between peer victimization, cyberbullying, and suicide in children and adolescents: A meta-analysis. JAMA Pediatrics, 168(5), 435–442. doi.org/10.1001/jamapediatrics.2013.4143
11. World Health Organization Regional Office for Europe (2024). A focus on adolescent social media use and gaming in Europe, central Asia and Canada — Health Behaviour in School-aged Children international report from the 2021/2022 survey, Volume 6. who.int
12. Office of the U.S. Surgeon General (2023). Social Media and Youth Mental Health: The U.S. Surgeon General's Advisory. U.S. Department of Health and Human Services. hhs.gov
13. Naslund, J. A., Bondre, A., Torous, J., & Aschbrenner, K. A. (2020). Social Media and Mental Health: Benefits, Risks, and Opportunities for Research and Practice. Journal of Technology in Behavioral Science, 5(3), 245–257. doi.org/10.1007/s41347-020-00134-x
14. Vuorre, M., & Przybylski, A. K. (2024). Global Well-Being and Mental Health in the Internet Age. Clinical Psychological Science, 12(5), 917–935. doi.org/10.1177/21677026231207791
15. Sigaud, L., Rausch, Z., McClean, A., & Haidt, J. (2026). Why Three Studies by Vuorre and Przybylski Should Not Be Used to Assess the Impact of Social Media on Adolescent Mental Health. Clinical Psychological Science. doi.org/10.1177/21677026261425910
16. Parry, D. A., Davidson, B. I., Sewall, C. J. R., Fisher, J. T., Mieczkowski, H., & Quintana, D. S. (2021). A systematic review and meta-analysis of discrepancies between logged and self-reported digital media use. Nature Human Behaviour, 5(11), 1535–1547. doi.org/10.1038/s41562-021-01117-5
17. Godard, R., & Holtzman, S. (2024). Are active and passive social media use related to mental health, wellbeing, and social support outcomes? A meta-analysis of 141 studies. Journal of Computer-Mediated Communication, 29(1), zmad055. doi.org/10.1093/jcmc/zmad055
18. Burnell, K., Meter, D. J., Andrade, F. C., Slocum, A. N., & George, M. J. (2025). The effects of social media restriction: Meta-analytic evidence from randomized controlled trials. SSM – Mental Health, 7, 100459. doi.org/10.1016/j.ssmmh.2025.100459
19. Maheux, A. J., Burnell, K., Maza, M. T., Fox, K. A., Telzer, E. H., & Prinstein, M. J. (2025). Annual Research Review: Adolescent social media use is not a monolith — toward the study of specific social media components and individual differences. Journal of Child Psychology and Psychiatry, 66(4), 440–459. doi.org/10.1111/jcpp.14085