How Much of Teen Mental Health Is Actually About Screens?

Key Takeaways

• Across 355,358 adolescents in three large datasets, the association between digital technology use and wellbeing was negative and very small, accounting for at most 0.4% of the variation between teenagers.
• The same paper found more than 600 million defensible ways to analyse that data. Depending on which choices an analyst makes, the same numbers yield a harmful association, a protective one, or none, which is why confident headlines exist in both directions.
• A 2025 analysis of two longitudinal cohorts tried to rank causes rather than measure one. Peer relationships had a strong direct effect on how mental health developed. Physical activity had a small one. Screen time was very small to negligible.
• All of this describes averages across populations. It says nothing about whether a particular child is being harmed by a particular app, and it is not a reason to ignore what you can see in your own house.
• The practical read is that screen time is a weak lever on this outcome and friendships are a strong one, so a plan built entirely around the phone is aimed at the smaller of the two.

What the largest measurement found

The most-cited attempt to size this association came out of Oxford in 2019, published in Nature Human Behaviour. Rather than run one analysis, the authors ran the analysis every reasonable way it could be run across three large datasets covering 355,358 adolescents in the US and UK.

The association between technology use and adolescent wellbeing was negative, and it was tiny. Digital technology use accounted for at most 0.4% of the variation in wellbeing between teenagers. The authors compared that with other things measured in the same datasets, and screen use sat alongside effects of a similar size to eating potatoes.

That comparison got the attention, and it is the weakest part of the paper to lean on, because an effect being small on average is not the same as it being unimportant for everyone. The stronger contribution is the method.

Why the same data produces opposite headlines

The Oxford team catalogued the analytical choices available to anyone studying this question: which wellbeing measure to use, which technology measure, which control variables, which subgroup. Combined, those choices produced more than 600 million defensible specifications of the same question on the same data.

Running all of them showed that the answer you get depends heavily on which you pick. Some specifications produce a harmful association, some produce a protective one, and many produce nothing. A researcher who runs one analysis and publishes it is not necessarily doing anything wrong, and is also not telling you where their result sits in that distribution.

This is the useful thing to carry into any headline about screens and teenagers. A published finding is one reading out of the many the same data allows, and the number of alternatives is the part that rarely travels with the headline.

What outranks screens when you try to rank causes

Measuring an association is a different job from ranking causes. A 2025 study in JMIR Public Health and Surveillance attempted the second, using two longitudinal cohorts: the British Millennium Cohort, with 8,599 participants, and the German KiGGS survey, with 1,212.

Instead of testing one predictor, it put many into a machine-learning model and used permutation and partial-dependence analysis within a causal inference framework to estimate each one's direct effect while holding the others constant. Three findings came out in order of size.

Peer relationships had a strong direct effect on how mental health developed, in both cohorts. Physical activity had a small direct effect, most visible among adolescents starting from very low activity. Screen time had a very small to negligible direct effect.

The authors are careful about what that supports. Causal inference from observational data has known limits, the two cohorts differed in design and only partly overlapped in what they measured, and factors like inequality and educational pressure were not in the models at all.

What this does not license

It does not license telling a worried parent that screens are fine. Averages across tens of thousands of adolescents can hide a subgroup being harmed badly, and none of this addresses a specific child, a specific app, or a specific pattern of use like sustained bullying or an eating-disorder feed. A small average effect and a serious individual harm can both be true in the same dataset.

It also does not settle the causal question. The 2019 work is explicitly correlational. The 2025 work applies causal methods to observational data, which is a step past correlation and short of an experiment.

What it does license is scepticism about confident claims in either direction, especially claims that a timeline is an explanation. Several things changed in adolescence over the same period, and picking one of them requires evidence rather than chronology.

What to do with it

If you are deciding where to put effort, the ranking is the useful part. A teenager with a phone and a strong friendship group is in a different position from one with a phone and no reliable friends, and the second situation is the one these datasets associate with how mental health actually develops.

The phone still matters. What to do first is ask who your teenager sees in an ordinary week, and whether the phone is displacing that or is how they arrange it. If the phone is how the friendships happen, taking it away removes the friendships. If the phone is what fills the time where friendships used to be, the problem was never the screen time figure.

For the narrower questions, whether monitoring is worth its cost and what school phone bans have actually been shown to do both have better evidence behind them than the broad causal question does.