Skip to content
FourthSpace

Phones, social media and teen mental health: a specification curve

Is heavy social media use tied to worse mental health among U.S. high school students? How sure can we be, and how big is the gap next to other things in a teenager's life?

Analysis by Andrew Ly (opens in a new tab) · Updated

Key finding

33.3%

High school students with persistent sadness or hopelessness

The association is consistent. 485 of 540 models (90%) put heavy users at higher risk.

Difference in prevalence (percentage points)

  • 95% interval includes zero
  • 95% interval excludes zero
  • Median

spec_curve_results.csv

Findings

This topic is contested. Jonathan Haidt argues that phone-based childhood drove the teen mental health decline. Candice Odgers, Amy Orben and Andrew Przybylski argue the evidence is small, mixed and mostly correlational. A specification curve is the method Orben and Przybylski used to show how results shift with analyst choices. Here it is run on the newest national data, including the 2025 YRBS released by CDC on September 21, 2026.

  1. The association is consistent.

    485 of 540 models (90%) put heavy users at higher risk. 300 (56%) have a 95% interval above zero. None has a 95% interval below zero. The median gap is 5.0 percentage points. Source: YRBS 2023 and 2025, data/clean/spec_curve_results.csv.

  2. It is larger for girls.

    The median gap is 7.6 points for girls and 3.0 for boys. In 2025, girls who use social media more than once an hour reported persistent sadness 11.2 points more often than other girls (95% CI 7.4 to 14.9), after sex, grade and race controls.

  3. It depends on the outcome.

    Every model for sadness and for poor mental health is positive. For seriously considering suicide, the median is 1.4 points and only 24% of models clear zero.

  4. Bullying absorbs part of it.

    Adding bullying as a control drops the median from about 5.5 to 3.7 points. Bullying may sit on the path between social media and sadness, so this control could also remove part of a real effect. The data cannot settle which.

  5. Other experiences show far larger gaps.

    In 2025, being bullied electronically goes with a 31.3-point higher rate of persistent sadness. Being bullied at school goes with 27.8 points. Not feeling close to people at school goes with 17.3 points. Using social media more than once an hour goes with 8.3 points. All four are adjusted for sex, grade and race, and all are associations.

  6. Sadness fell while heavy use held steady.

    From 2023 to 2025, persistent sadness fell from 40% to 33%. Use about hourly or more held at 37% and 36%. Use several times a day rose from 77% to 80%. A change in the mix of social media use explains +0.2 points of the 7.1-point drop. The other 7.3 points of decline happened inside every frequency group.

  7. School connection improved.

    The share who do not feel close to people at school fell from 45% to 37%. CDC's 2025 summary report left this item out, so this comes straight from the public microdata.

  8. Teens remain heavily online.

    40% of U.S. teens say they are online almost constantly (Pew, 2025), up from 24% in 2014-15 and down from 46% in 2022 through 2024. TikTok is the app the most teens use almost constantly (21% of all teens).

Options

  1. Option A. Lead with the phones narrative

    Use Haidt's framing on the landing page and in the deck. It is familiar to readers. It also invites pushback the data here does not fully back.

  2. Option B. Lead with connection

    My pick

    Say heavy use goes with more sadness, then show that belonging and safety at school go with much larger gaps, and that sadness fell as school connection improved. FourthSpace sells places to belong. This survives critics on both sides.

  3. Option C. Position FourthSpace as light-use tech

    Cite the WHR 2026 finding that young people using social media under an hour a day report the highest wellbeing. Pitch the app as one you open for a minute to go somewhere.

  4. Option D. Keep teen data out of public materials

    FourthSpace serves adults. Keep this project as portfolio evidence of method and stay out of the teen debate in marketing.

  5. Option E. Go local next

    Request Sacramento County results from the California Healthy Kids Survey (WestEd/CDE) to see whether the school connection pattern holds for local teens.

Default pick: B. It is the reading the data supports best. It holds up without settling the causal debate, and it gives FourthSpace a positive message about what helps. C pairs well with B in the product story.

Method

  1. 01_yrbs_load.py reads both YRBS fixed-width files using the column positions in CDC's SAS input programs. Question numbers moved between 2023 and 2025, so each year has its own map. Response codes match across years.
  2. Validation. Weighted prevalences reproduce CDC's published figures. 2023: sadness 39.7% (CDC 40%), considered suicide 20.4% (20%), frequent social media 77.0% (77.0%). 2025: sadness 33.3% (33%), girls 43.4% (43%), boys 23.9% (24%), considered suicide 14.5% (14%), hourly or more social media 35.8% (36%).
  3. 02_spec_curve.py runs every combination of five analyst choices. That gives 6 definitions of heavy use, 3 outcomes, 5 control sets, 3 samples and 2 survey years, for 540 models. Each model is a survey-weighted linear probability model. The estimate is the adjusted gap in prevalence, in percentage points.
  4. 03_benchmarks_shiftshare.py puts the gap in context. It runs the same model for other exposures in the same survey, then a model with all of them together. It also splits the 2023 to 2025 drop in sadness into the part from a change in how often students use social media and the part that happened inside each frequency group.
  5. 04_pew_teens.py pulls each trend row from the Pew topline PDF and stores the exact source line next to each number.
  6. 05_charts.py draws every chart as PNG and interactive HTML.

Tools

Python 3.13, pandas, NumPy, pdfplumber (Pew PDF parsing), matplotlib (PNG charts), Plotly (interactive HTML), pyarrow (parquet). Survey variance is computed by a small custom module (scripts/survey.py) that does Taylor-linearized, design-based standard errors with strata and PSUs, the same approach CDC uses. Data was downloaded with curl on a Mac (fetch_data.sh at the repo root).

Caveats

  1. Correlation only. No model here identifies cause.
  2. YRBS measures frequency of use. It has no item on hours, content or apps. Frequency is a rough stand-in for the kinds of use researchers think matter.
  3. Students who do not use social media are a small group (7% in 2025). Their estimates have wide intervals.
  4. The 2025 YRBS sample is about half the size of 2023 (10,710 vs 20,103). CDC also changed sampling and mode in 2023. Two years is a short trend.
  5. The linear probability model gives gaps in percentage points. CDC's own report uses prevalence ratios from logistic models. The direction matches. The scale differs.
  6. Some questions were asked of a subset of students (sports team, school connection), so models that use them have smaller samples.
  7. The suicide item is a serious public health measure. It stays out of public-facing FourthSpace materials.

Downloads

These files come from US federal agencies and are in the public domain. Every other source has its own terms, so it is linked at the publisher under Sources.

Sources

  1. CDC National Youth Risk Behavior Survey, public-use microdata (n=20,103)

    Year
    2023
    Geography
    U.S. high school students

    https://www.cdc.gov/yrbs/data/national-yrbs-datasets-documentation.html (opens in a new tab)

    Public domain

  2. CDC National Youth Risk Behavior Survey, public-use microdata (n=10,710)

    Year
    2025
    Geography
    U.S. high school students

    https://www.cdc.gov/yrbs/files/2025/XXH2025_YRBS_Data.dat (opens in a new tab)

    Public domain

  3. Pew Research Center, Teens & Technology topline (N=1,458, ages 13-17, MOE ±3.3 pts)

    Year
    Sep-Oct 2025, trend back to 2014-15
    Geography
    U.S. teens

    https://www.pewresearch.org/internet/ (opens in a new tab) (topline PDF in data/raw)

    Publisher's terms

  4. World Happiness Report 2026, social media chapter findings (cited, no microdata)

    Year
    2026
    Geography
    Global

    https://www.ox.ac.uk/news/2026-03-19-world-happiness-report-2026-shows-complex-global-picture-social-media-and-happiness (opens in a new tab)

    Publisher's terms

  5. CDC MMWR, Young et al., frequent social media use and YRBS 2023 outcomes (used to validate the 77% figure)

    Year
    2024
    Geography
    U.S.

    https://www.cdc.gov/mmwr/volumes/73/su/su7304a3.htm (opens in a new tab)

    Public domain

Debate sources are listed in the debate box below.

The strongest evidence on each side

The case that phones and social media are a major cause

Jonathan Haidt's The Anxious Generation (2024, https://www.anxiousgeneration.com (opens in a new tab)) argues that a shift to phone-based childhood between about 2010 and 2015 drove rising teen anxiety and depression. The evidence offered includes the timing across many countries, girls being hit hardest, dose-response patterns like the one in this project, and some experiments where people who cut back report feeling better. The World Happiness Report 2026 adds that young people's life evaluations fell most in English-speaking countries, with heavy use tied to much lower wellbeing.

The case that the evidence is weak or misread

Amy Orben and Andrew Przybylski ran specification curves on three large datasets including YRBS and found digital technology use explained at most 0.4% of the variation in adolescent wellbeing (Nature Human Behaviour, 2019, https://doi.org/10.1038/s41562-018-0506-1 (opens in a new tab)). Candice Odgers reviewed Haidt's book in Nature and argued that the research does not show social media causes mental illness, and that the focus may pull attention from known causes such as poverty and adversity (2024, https://doi.org/10.1038/d41586-024-00902-2 (opens in a new tab)).

What this project's data can and cannot say

Every number here is cross-sectional. It shows which students report more sadness. It cannot show what caused what. Sad students may use social media more. Bullying, family stress or sleep may drive both. The findings fit both camps in part. The association is real, positive and larger for girls, as Haidt says. It is also modest next to bullying and school connection, and the 2023 to 2025 drop in sadness happened with no fall in heavy use, which supports the critics' point that other forces move teen mental health.

Interactive charts

  • Bullying and weak school ties track with sadness more than heavy social media use

  • Sadness is highest among students who use social media more than once an hour

  • Heavy social media use lines up with worse teen mental health in 9 of 10 models

  • About 4 in 10 teens say they are online almost constantly

  • Teen sadness fell from 2023 to 2025 while hourly social media use held steady

Keep reading

What connection is worth: loneliness, health and Sacramento

How strongly is social connection tied to health and survival? How common is loneliness in the U.S. and in Sacramento County, and how many Sacramento adults does that represent?

Next study
Every major meta-analysis finds higher mortality among less connected adults

Be first through the door.

We're opening in Sacramento first. Join the list and we'll email you when we open.

Fields marked * are required.

Places you'd go (optional)
How we use your details

Study02/07Join the waitlist