The short answer

The evidence cited here is not enough for “phones caused the crisis”, nor for “it does not matter”.

Persistent sadness has risen in the US data: the share of high-schoolers who say they felt so sad or hopeless almost every day for two weeks that they stopped usual activities ranged between 33.9% and 36.7% for girls in 1999–2011 and reached 52.6% in 2023; for boys, from 21.5% in 2011 to 27.7% in 2023 (CDC). The rise is real in these data. Two lines rising together do not show a cause, however. Jonathan Haidt, Zach Rausch and Jean Twenge tentatively conclude that there is a causal link; Candice Odgers writes, in a Nature book review, that researchers have found “a mix of no, small and mixed associations”; Orben and Przybylski judged the link “too small to warrant policy change”.

The evidence splits in three: what surveys show (small links on average, larger at heavy use), what experiments show (small effects in the meta-analysis of students and adults; in minors, little evidence, including a moderate two-week effect of reducing screens in general) and what school bans show (no randomised trial; mixed results).

−0.035median of the 372 analyses of one survey; 356 are negative, 323 differ significantly from zero
49.8% vs 32.0%of girls say they felt sad or hopeless: those using devices five hours or more a day, against those under an hour (boys: 30.9% vs 16.5%); 2007–2015 pooled, and the devices question widened in 2013
23.4 yearsaverage age of participants in the 32 articles on social-media reduction experiments pooled (5,544 people), all students or adults

In the US survey analysed here the link is small on average and, in almost every analysis, of the same sign: more screen time, less favourable answers. Other syntheses give from nearly zero up to r = .10–.15. Cause cannot be established from surveys, and experiments on adolescents are few and small. These findings do not establish whether phones caused the rise in adolescents' sadness, nor their long-term effects.

The trend both camps rely on

The share of US high-schoolers answering “yes” to the persistent-sadness question, by sex, in each year of the national survey (CDC). The line stayed nearly flat until 2011 and rose after. The 2021 survey was run in the autumn, not the spring as usual, because of the pandemic (CDC).

Year	All	Girls	Boys
1999	28.3	35.7	21.0
2001	28.3	34.5	21.6
2003	28.6	35.5	21.9
2005	28.5	36.7	20.4
2007	28.5	35.8	21.2
2009	26.1	33.9	19.1
2011	28.5	35.9	21.5
2013	29.9	39.1	20.8
2015	29.9	39.8	20.3
2017	31.5	41.1	21.4
2019	36.7	46.6	26.8
2021	42.3	56.6	28.6
2023	39.7	52.6	27.7

One survey, 372 analyses

A 2019 study (Orben and Przybylski, Nature Human Behaviour) ran every defensible analysis on three large datasets and concluded that screens explain at most 0.4% of the variation in well-being, “too small to warrant policy change”. The set with the fewest analyses, the US survey YRBS (75,083 students, 2007–2015), has 372 analyses. Each dot below is one.

β is the standardised slope of the link between time on screens and the answers to five survey questions (without a control it is the correlation itself): 0 means no link, and negative values mean that more screen time goes with less favourable answers. The 372 analyses were redone with independent code on the CDC files and match the authors' published table row by row.

The five survey questions: felt sad or hopeless almost every day for two weeks; seriously considered suicide; made a plan; attempted; needed medical care after an attempt. They are survey questions, and the figures are link coefficients, not counts of cases.

If the subject touches you personally: Telefonul Copilului (the Romanian child helpline), 116 111 (Monday–Friday, 10:00–20:00, free), and 112 in an emergency; details below, under “Where to ask for help”.

The critique: the authors “made six analytical decisions that resulted in lower effect sizes”, and their conclusions contrast with “the practically important differences” found in other analyses of the same data, “especially for social media use among girls” (Twenge et al., 2020).

When the filters move, the result moves too: that is why all the analyses should be seen, not a chosen one. The critics (Twenge, Haidt, Joiner and Campbell, 2020) say the authors' analytical decisions led to smaller effect sizes; the authors (Orben and Przybylski, 2020) reply that the critics “incorrectly” concluded that the small associations are the product of those decisions, and add further analyses. The panel shows how far the result moves when what counts as “screens”, which questions count, girls and boys separately and the survey year change.

Averages by category

Median β of the 372 analyses, by sex and by screen measure (YRBS 2007–2015, unweighted). For television, girls have a coefficient near zero; for devices, the girls' coefficient is slightly larger than the boys' (a small difference); for the mean of the two screens and across all 372 analyses it is larger for boys.

Screen	All	Girls	Boys
All 372	−0.035	−0.033	−0.059
Television	−0.012	0.000	−0.028
Devices	−0.071	−0.084	−0.075
Mean of the two	−0.049	−0.056	−0.067
Negative	356	310	372
Significantly negative	323	261	364

By single question, with devices and no control (β, girls / boys): sad or hopeless −0.100 / −0.084; suicidal thoughts −0.087 / −0.077; plan −0.085 / −0.074; attempt −0.064 / −0.054; medical care after an attempt −0.028 / −0.036.

Share answering “yes” to sadness, by hours a day

One coefficient for everyone flattens the shape of the link. When students are grouped by hours on devices on an average school day, the share answering “yes” to the persistent-sadness question does not rise evenly: for girls it is nearly the same from “none” to two hours, then climbs to 49.8% at five hours or more; for boys it goes from 16.5% (under an hour) to 30.9%. Averages are weighted with the survey weights (2007–2015 pooled).

Hours a day	Girls	n	Boys	n
none	33.8 (32.4–35.2)	7,771	19.2 (17.6–20.9)	4,708
under 1 hour	32.0 (29.9–34.1)	6,771	16.5 (15.2–17.8)	6,585
1 hour	32.2 (30.3–34.2)	5,145	17.6 (16.4–18.9)	5,333
2 hours	34.8 (33.1–36.6)	5,435	19.6 (18.4–20.8)	6,327
3 hours	39.7 (37.3–42.1)	3,948	20.3 (18.8–21.7)	4,874
4 hours	43.9 (41.1–46.7)	2,312	22.5 (20.1–24.9)	2,829
5 hours or more	49.8 (47.5–52.0)	5,045	30.9 (29.1–32.7)	5,326

The devices question changed meaning across years: in 2007–2011 it was about video games or computer; from 2013 it explicitly counts the phone, tablets and social media. Years are therefore not directly comparable. For television, the differences between groups are smaller (9.1 percentage points for girls, 9.0 for boys, against 17.8 and 14.4 for devices). The shape is not strictly rising: for boys, the share at “none” is higher than at “under 1 hour”.

What does not follow: it does not follow that hours on devices cause sadness. Students who spend five hours or more a day on devices may differ in many other ways, and sadness may lead to more screen time. The data are from a single moment (each student answered once), so they do not show what came first.

Social media, asked directly in 2023

The 2007–2015 survey had no question on social media separate from games and computer. The 2023 one does: “How often do you use social media?” Across its eight answers, the share saying they felt sad or hopeless is, for girls, 52.1% at “several times a day” and 58.7% at “more than once an hour”; for boys, 28.4% and 32.8%. For those who do not use social media at all, the confidence intervals are very wide (32.6–60.3% for girls, 375 girls; 11.4–35.5% for boys, 672 boys), so they show no certain difference; for girls, the share at “does not use” (46.5%) is higher than at “about once a day”, again with a wide interval.

The table behind the chart

How often	Girls	n	Boys	n
does not use	46.5 (32.6–60.3)	375	23.4 (11.4–35.5)	672
a few times a month	38.6 (20.6–56.7)	121	31.2 (11.4–51.0)	269
about once a week	45.2 (22.0–68.4)	97	28.2 (20.1–36.4)	121
a few times a week	47.7 (39.4–56.1)	309	22.6 (16.8–28.4)	386
about once a day	40.0 (32.5–47.4)	328	21.1 (14.6–27.6)	565
several times a day	52.1 (48.3–55.9)	3,039	28.4 (26.2–30.6)	2,791
about once an hour	56.4 (48.9–63.9)	542	24.0 (18.0–30.0)	622
more than once an hour	58.7 (55.1–62.3)	2,628	32.8 (29.7–35.9)	2,119

Share of those who answered: among girls 42.0% at “several times a day” and 39.7% at “about once an hour or more”; among boys 38.5% and 34.4%.

Limits: for 24.4% of students there is no recorded answer to the social-media question; the eight levels leave few students in the first three (wide intervals); a single year; a link, not a cause. The overall sadness shares for 2023 (girls 52.6%, boys 27.7%) match CDC's published figures.

What the experiments show

Cause is most directly tested when participants are assigned at random to use social media less or more; natural experiments can also support causal conclusions, with extra assumptions. Most experiments of this kind were done on adults and students.

Study	Who, how long	What was found	The authors' conclusion
Allcott et al. 2020	US Facebook users, four weeks before the 2018 election; account deactivation	Well-being: +0.09 standard deviations (overall index); in the NBER working paper, “small but significant” improvements, and the daily text-message measure is positive but not significant	“Deactivating … increased subjective well-being” (AER abstract)
Lambert et al. 2022	154 adults (mean age 29.6), one week without social media	Well-being: difference 4.9 [3.0; 6.8]; depression −2.2 [−3.3; −1.1]; anxiety −1.7 [−2.8; −0.6]	One week without social media led to significant improvements in well-being, depression and anxiety; the authors call for longer-term studies and clinical populations
Hunt et al. 2018	143 University of Pennsylvania undergraduates; 10 minutes per platform per day, three weeks	Less loneliness and depression in the limited-use group; both groups showed decreases in anxiety	“Limiting … to approximately 30 minutes per day may lead to significant improvement in well-being”
Burnell et al. 2025	Meta-analysis: 32 articles, 5,544 people, 91 effect sizes; “all studies included college student or adult samples”	ḡ = 0.17 [0.08; 0.27]; restricting use significantly improved subjective well-being; samples 70% female	Only “weak support for the effectiveness of restricting” social media
Braghieri, Levy & Makarin 2022	US college students, more than 430,000 survey responses; staggered arrival of Facebook on campuses (natural experiment)	Facebook's arrival at a college had “a negative impact on student mental health”	The authors link the effect to unfavourable social comparisons
Sullivan, Cairney & Henderson 2026	Preprint, not peer-reviewed: 82 UK adolescents aged 13–18; 21 days without smartphone and social media, with or without a basic phone	Tentative improvements in sleep, depressive symptoms, stress and working memory at 21 days; not maintained at two months	A pilot study of acceptability; the authors call for larger studies
Schmidt-Persson et al. 2024	89 Danish families, 181 children and adolescents (mean ages 8.6 and 9.5); screens in general (not only social media), two weeks	Total difficulties: −1.67 [−2.68; −0.67], d = 0.53; internalising symptoms −1.03 [−1.76; −0.29]	Randomised by family; a two-week effect, durability unknown

Experiments with minors are few and small: one published randomised trial (Schmidt-Persson et al.) cut screen time in general, not only social media, with participants averaging about 9 years old; a pilot study of 82 adolescents, not yet published, found improvements that were gone at two months. No peer-reviewed experiment in which adolescents reduced social media was found; that is a search with no result, not proof that none exists.

The review by supporters of the causal thesis (Haidt, Rausch and Twenge) counts the experiments that indicate harmful effects and those that do not; the document says of itself: “We are not unbiased”. The same document cites critics: in one experiment (Hunt), the reduction moved only 2 of 7 outcomes examined, and after correction for multiple comparisons one stayed significant (Patrick Markey). The restriction interventions in the table above last from one week to four.

Studies over time on adolescents

Two cohorts follow the same children: in a US sample (11,876 children, aged 9–10 at the start), social-media use above a child's own average was associated with more depressive symptoms the next year (β 0.07 between years 1 and 2, 0.09 between years 2 and 3), depressive symptoms were not associated with later use, and differences between children in social-media time were not associated with symptoms (β −0.01); in a Melbourne sample (1,195 adolescents aged 12–18), more than two hours a day against under one was associated with a small rise in the risk of high depressive symptoms the next year (6.3 per 100 [2.7; 9.9]; for girls in early adolescence 10.8 [2.7; 18.9]). An older study (Orben, Dienlin and Przybylski, 2019) found, in another sample, small associations in both directions, differing by sex and by analytical choices.

The average hides differences between people

In a study of 63 Dutch adolescents (mean age 15.12), followed six times a day for a week, 44% felt neither better nor worse after passive social-media use, 46% felt better and 10% felt worse. In a study of 84,011 participants aged 10–80, longitudinal analyses of 17,409 people aged 10–21 found “windows of sensitivity”, when more estimated social-media use predicts a drop in life satisfaction a year later: ages 11–13 and 19 for girls, 14–15 and 19 for boys. Decreases in life satisfaction also predicted later increases in estimated social-media use, regardless of age or sex. Both studies are observational.

The two camps, side by side

The table compares selected arguments from both camps; quotation marks mark direct quotations.

The question	If screens do harm (Haidt, Twenge)	If it is “not proven” (Odgers, Orben, Przybylski, Ferguson)
How large is the link?	The collaborative review, written by supporters, says researchers who reached opposite conclusions in 2019 now agree on a link of r = .10 to .15 for all teens, far from zero and larger for girls; it cites a review by Orben (2020): “about r = − 0.15 to r = − 0.10”.	The 2019 study: the medians of the three datasets run from −0.005 to −0.035 (at most 0.4% of variation), and the size depends on choices. A meta-analysis of 226 papers, mostly of young adults (preprint), gives r = 0.01 overall but larger for anxiety (0.13) and depression (0.12). Ferguson et al. (2022, 37 effect sizes from 33 studies): r = 0.052, concluding that screen media “plays little role”. A review of 25 reviews (Valkenburg et al., 2022) found that most described the associations as “weak” or “inconsistent”, and a few as “substantial” and “deleterious”.
Can it be the cause?	Reduction experiments report small improvements, and one natural experiment links Facebook's arrival to worse mental health in students; in one cohort of children, increased use is associated with symptoms the next year (Nagata 2025).	The social-media restriction experiments in Burnell's meta-analysis use student or adult samples; Odgers stresses that most data are correlational and that links over time also point the other way (Odgers 2024).
Who matters?	“Practically important differences”, “especially for social media use among girls” (Twenge et al., 2020).	In the YRBS data analysed above, the girl–boy gap depends on what is measured: the coefficient is larger for girls for devices and larger for boys for television.
What next?	Solutions proposed by Haidt, as Odgers cites them: minimum ages for platforms and phone bans (she judges them unlikely to be effective).	Odgers: “two things can be independently true”: there is no evidence that platforms are “rewiring” children's brains or driving an epidemic of mental illness, yet “considerable reforms” of the platforms are required; age limits and phone bans are, in her view, “unlikely to be effective in practice” or could “backfire”; separately, a generation in crisis needs help.

Odgers (Nature, 2024): “Most data are correlative.” Where links over time are found, they suggest, in her words, that “young people who already have mental-health problems” use such platforms more. Cohorts published after her review report prospective associations in the other direction: use above a child's own average was associated with later symptoms (Nagata 2025), and higher daily-use categories with symptoms the next year (Vijayakumar 2026).

The table combines attributed quotations, summaries and the analysis of the data above. The collaborative review by Haidt, Rausch and Twenge acknowledges its stance: “We are not unbiased”.

Banning phones in school

This is a different question from social media: if a school restricts phones, do grades or pupils' state change? There is no randomised trial. What exists are comparisons between schools.

Study	What was compared	What was found	What it does not show
Beland & Murphy 2016	English schools with a ban, before and after; 91 schools answered the survey (21% of those targeted)	Test scores: +6.41% of a standard deviation (significant only at p < 0.1); pupils in the lowest quarter of prior achievement: +14.23%; top achievers: no effect; no gains where a ban is not widely complied with	Nothing about mood; numbers are from the working-paper version of the article
Goodyear et al. 2025	30 English schools (20 restrictive, 10 permissive), 1,227 pupils aged 12–15; cross-sectional	Well-being: difference −0.48 [−2.05; 1.06]; phone time during school time 0.67 hours less; total phone time, no difference	The authors: no evidence that school restrictions are associated with less total phone use or better well-being; the data do not support bans “in their current form”
Abrahamsson 2026	Norwegian middle schools, before and after a ban	No average effect on education or mental health; for girls, use of mental-health care fell and grades rose; bullying fell for both sexes	The figures were not read in the abstract
Kohnstamm Instituut & Oberon 2025	National evaluation in the Netherlands (2025): 271 secondary schools with a policy, schools' answers	74.9% report a positive effect on concentration, 58.7% on pupils' well-being and school climate, 27.7% on performance; none reports a negative effect on concentration	Schools' perceptions, not measured outcomes
Böttger & Zierer 2024	Rapid review of five studies	Small effect: d = 0.162, larger for social well-being than for grades; the authors recommend bans accompanied by educational measures and regular evaluation	Only five studies
Campbell et al. 2024	Scoping review of 22 studies	“An absence of randomized controlled trials”; designs and measures very different, making reconciliation of findings challenging	Does not show an effect

In short: pupils in schools with restrictions use their phones less during school time; the effect on grades is small and clearer for weaker pupils (England); an effect on well-being does not appear in the English comparison, and in the Norwegian data girls' use of mental-health care falls and bullying falls for both sexes. A report prepared for the European Commission (ENESET, 2026) finds that “evidence on the impact of school bans remains limited”.

The number of education systems with a national ban has grown: under a quarter of countries in 2023, 114 systems (58% of countries) in 2026, according to UNESCO (Global Education Monitoring Report).

What the rules say

Country	Rule	Source
Romania	Law no. 198/2023, art. 106: phones banned during lessons in pre-school, primary and lower secondary, except for educational use and spaces allowed by the school's own rules; upper-secondary schools may ban them on the whole premises by internal rule; equipment that pupils with special educational needs are authorised to use is exempt	the law (Official Gazette 613/2023)
France	A 2018 law bans phones in primary and lower-secondary schools, breaks included; from 2025 the “digital pause” (pupils hand in their phone on arrival) was made general in lower secondary, up to age 15	ENESET report for the European Commission, 2026
Netherlands	A 2024 national agreement, not binding by law but strongly urged: no phones during lessons, in secondary from January 2024 and in primary after the summer of the same year; the principle is “no, unless” (for example, when needed for the lesson or for medical reasons)	Dutch government, 12 July 2024; the ENESET report
Italy	Circular of July 2024: a ban for the whole day, breaks and educational use included, for ages 3–14; extended in June 2025 to upper secondary; exceptions for pupils with individual education plans	ENESET report for the European Commission, 2026
Denmark	Political agreement of 30 September 2025: primary and lower-secondary schools (ages 6–16) without phones during school time, with local policies, from the 2026/27 school year; legislation is being prepared	ENESET report for the European Commission, 2026; Eurydice
Hungary	Decree 245/2024, in force from 1 September 2024: pupils hand in their phone at the start of the day, it is kept locked and returned after the last lesson; educational and medical exceptions	ENESET report for the European Commission, 2026; Eurydice

In 2026 the Senate passed a bill by the AUR party on the use of phones in upper-secondary schools (59 votes in favour, of 96 senators present), and the education minister, Mihai Dimian, said: “At the moment the law is not being respected” (press; the bill passed only the Senate and is not law; the bills' official texts were not checked at source). Edupedu Wall-Street.ro

Romania in the data

Two international sources say something about Romanian adolescents. Neither analyses here the link between screens and Romanian pupils' mood.

The WHO's HBSC 2021–2022 study (almost 280,000 adolescents aged 11, 13 and 15, 44 countries and regions): Romania has the highest share of “problematic social media use”, 22% against an average of 11%, followed by Malta (18%) and Bulgaria (17%); Romania also had the largest increase. “Problematic” means the adolescent answered “yes” to six or more of nine statements about symptoms (symptoms include difficulty controlling use, distress when use is restricted and preoccupation with social media when not online); it is a self-report scale of nine yes-or-no questions. The report notes that problematic use is associated with lower mental, social and school well-being. At age 13 the share is 21% for boys and 28% for girls.

In PISA 2022, 68.6% of Romanian 15-year-olds said they are distracted in maths lessons by other students' devices in some, most or every lesson, against 59.3% for the OECD average. Coverage of the 15-year-old population in Romania is between 50% and 75%, which the OECD flags with the symbol †.

No analysis linking screen time to pupils' mood in Romania was found from primary sources. The raw HBSC 2021–2022 data were under embargo until October 2026.

Method

Data: the CDC's national YRBS files for 2007, 2009, 2011, 2013 and 2015 (75,083 students, of whom 37,402 girls and 37,412 boys with sex recorded, as in the study) and 2023 (20,103 students). Items change number from year to year, so they were matched by question text.
The 372 analyses: three screen measures (television, devices, their mean) × 31 combinations of the five questions × two combination rules × with or without a control for race/ethnicity (white non-Hispanic or not). Each analysis is a linear regression on standardised variables; β is the standardised slope.
Check: the authors' table of 372 analyses (OSF, CC BY 4.0) was compared row by row with the output of the independent code: maximum difference in β 6e-14 (practically zero), same n in every row, median −0.035 (in the study: −0.035), 356 negative and 323 significantly negative (in the study: 356 and 323). The authors' MTF and MCS tables are shown without recomputation; their medians match those in the study's text.
The “problem on every chosen question” rule: the published table is reproduced when a respondent falls outside the category only by reporting a problem on every chosen question; the study's text describes the rule as “yes to one or more questions”. The page reproduces the published table.
The own extension (by sex and by year) runs the same 372 analyses on subsamples; it is not in the original study.
Weights: the 372 analyses are unweighted, as in the original study. The percentages in the “hours a day” and “social media” charts are weighted with the survey weights, with 95% confidence intervals from cluster variance (stratum and primary sampling unit). The overall sadness shares match CDC's published ones (2013: 39.05% for girls in both sources; 2023: 52.6% and 27.7%).
The figures on experiments, bans and Romania come from the abstracts or published texts of the cited sources.

Limits

A US survey of high-school students; answers are self-reported; each student answered once, so the order in time cannot be seen.
The devices question does not separate social media in 2007–2015; it widened in 2013 (it includes the phone and social media).
The only control in the YRBS table is race/ethnicity; other factors (sleep, family circumstances, income) are not controlled.
Persistent sadness is measured by a single question; the suicide questions are survey questions, and the result is not a count of cases.
MTF and MCS: the authors' tables are shown; without the raw data (access by registration) they cannot be redone.
Experiments with minors are few and small; the search made here found no peer-reviewed experiment specifically asking adolescents to reduce social media.
Nothing here is medical advice, and the data do not establish the phone's long-term effects on adolescents. Review: when the HBSC 2021–2022 data are released or by 3 January 2027.

Data and code

The archive holds the computed results (CC BY 4.0), the code (MIT) and a script that downloads the original files (CDC, OSF), checks them by SHA-256 and redoes the 372 analyses. It does not hold the CDC files, nor the papers or abstracts, because the archive has no right to redistribute them on the sources' behalf.

To redo it locally:

pip install -r requirements.txt
cd code
python3 reproduce.py

The CDC blocks automated downloads: the script tries the CDC first, then a copy on the Internet Archive, and checks SHA-256; if both fail, it says which file to download by hand from the CDC page.

The computed results are under CC BY 4.0, the code under the MIT licence. The CDC files are published by the CDC with no access conditions on its download page; the authors' OSF tables are under CC BY 4.0.

Phones and teenagers' mood: what the public data show