Screen use explained at most 0.4% of the variation in adolescent well-being

When researchers applied a method designed to stop analysts cherry-picking their own results to three large datasets covering 355,358 adolescents, the association between digital technology use and well-being came out negative but small — explaining at most 0.4% of the variation in well-being [s1]. The authors' own reading was that effects of that size are too small to warrant policy change [s1].

That is not the same as saying there is no association. It is a statement about how big the association is, which is the part of this argument that usually goes missing.

The method matters more than the number

The problem the analysis set out to solve is specific. Evidence on this question comes largely from secondary analyses of large social datasets, which contain so many variables and observations that analytical flexibility alone can mark small effects as statistically significant — producing false positives and conflicting results [s1].

Specification curve analysis addresses this by running the enormous number of defensible analyses a dataset permits, and reporting the distribution of results rather than one chosen path. Applied across three large-scale social datasets with a combined n of 355,358, it returned a negative association that at its largest explained 0.4% of the variation in adolescent well-being [s1].

The same researchers then went at the measurement problem. Most results in this area rely on self-reported estimates of screen use, which the authors describe as inaccurate but popular [s2]. In a follow-up study they used time-use-diary measures instead, across three nationally representative datasets from Ireland, the United States and the United Kingdom, with N = 17,247 after data exclusions, and used both exploratory and confirmatory designs [s2].

The result: little evidence for substantial negative associations between digital-screen engagement and adolescent well-being — whether the engagement was measured throughout the day or specifically before bedtime [s2].

The bedtime clause is worth pausing on, because pre-sleep screen use is the version of this hypothesis with the most plausible mechanism behind it, and it is the version the time-use diaries also failed to support [s2].

The pairing of exploratory and confirmatory designs in that second study is not decoration either [s2]. An exploratory analysis searches a dataset for patterns; a confirmatory one commits to a prediction and tests it. Reporting both makes it possible to see whether a result found by looking survives being looked for. In this literature, where the same public datasets have been analysed repeatedly by different teams reaching different conclusions, that distinction does much of the work of separating a finding from an artefact of searching.

Where the association is not the same everywhere

A preregistered scoping review of research published between 2018 and 2020 examined 34 articles on adolescent social media use and depression, asking a question the rest of this literature had largely skipped: who was studied [s3]?

More than 70% of the studies examined Global North populations [s3]. And the association differed by region: the link between social media and depression was positive and significant in the Global North, but null and nonsignificant in the Global South [s3]. The review also found little evidence of diversity within regions, in either sampling choices or the reporting of participants' demographics [s3].

Given that most adolescents live in the Global South, the review's authors describe this as requiring urgent action, on the grounds that sample diversity is essential to whether findings generalise at all [s3].

That finding does two things at once. It undercuts any claim that the association is a universal property of adolescence and social media. And it raises the possibility that what is being measured in the Global North is contextual — something about a particular set of platforms, populations or circumstances rather than about screens as such.

What the small effect does and does not license

A small average association is compatible with a large effect in a subgroup. None of these three studies rules out that some adolescents are substantially harmed by some kinds of use; they measure population averages and, in the third case, regional differences [s1] [s2] [s3].

They are also all correlational. The specification curve analysis is explicitly framed as examining correlational evidence [s1], and the time-use-diary work, for all its measurement improvements, still observes rather than intervenes [s2]. Neither can establish direction: adolescents who feel worse may use their phones more.

And these are datasets from the late 2010s. Platforms, use patterns and the amount of public attention on this question have all changed since.

The claim that survives

The honest version is narrower than either side of the public argument. There is a negative association between adolescent digital technology use and well-being. It is small — at most 0.4% of the variance in the largest analysis [s1]. It does not strengthen when screen use is measured well rather than self-reported [s2]. And it does not replicate outside the Global North, where most adolescents in the world actually live [s3].

Those are findings about effect size and generalisability, not a verdict that phones are harmless. The most defensible thing to say is that the evidence has been asked to carry far more weight than its magnitude supports, and that the studies best designed to detect a large effect did not find one.