13 Features. 613K Students. One Pattern.

What if you could measure it?

Everyone says social media harms teens. But which features? Which design choices? We scored 13 specific, verifiable design features across 10 platforms for each year from 2011 to 2023.

Opacity (O-type)

Algorithmic feed
Autoplay video
Opaque recommendation
Hidden ranking signals

Avg R² = 0.549

Reactivity (R-type)

Infinite scroll
Push notifications
Real-time metrics
Streaks / daily hooks

Avg R² = 0.493

Coupling (α-type)

Beauty / AR filters
Social comparison visible
Identity persistence
Disappearing content
Default-public minors

Avg R² = 0.375

Every feature is a checkable fact. Does Instagram have infinite scroll? Yes or no. No expert judgment. No subjective scoring. Any opposing expert can verify the codings from app changelogs and archived interfaces.

Which features predict harm?

Average R² across all six YRBS mental health outcomes. Longer bar = stronger predictor of teen depression, sadness, and suicidality.

opaque_recommendation

0.852

real_time_metrics

0.802

social_comparison

0.802

autoplay_video

0.780

algorithmic_feed

0.765

hidden_ranking_signals

0.549

infinite_scroll

0.523

push_notifications

0.493

identity_persistence

0.456

streaks

0.420

beauty_filters

0.380

disappearing_content

0.350

default_public_minors

0.320

R² = 0.938

opaque_recommendation alone explains 93.8% of the variance in female teen persistent sadness (N=7 waves, wide CI).
One feature. One design choice. One number a jury will remember.

What the data says

CDC Youth Risk Behavior Survey, 2011–2023. Seven biennial waves. 13,000–20,100 students per wave.

R² = 0.80

Feature-weighted exposure explains 80% of the variance in teen persistent sadness. Raw social media adoption explains 70%. The features matter more than whether you use the app.

5.6×

Girls are 5.6× more affected than boys in 91% of 47 countries. Replicated on independent OECD PISA data from 613,744 students across 80 countries (p < 0.000001).

+319%

Feature exposure increased 319% from 2011 to 2023 (14.2 → 59.5). Raw adoption only increased 137%. The platforms got more manipulative, not just more popular.

R² = 0.055

Electronic bullying: ZERO signal (p = 0.61). The methodology does not correlate with everything. It identifies the specific features driving specific harms. This is the negative control.

The 2016 inflection

Instagram switched from chronological to algorithmic feed in mid-2016. This is the single largest feature exposure jump in the dataset.

+17.1

Feature exposure jump (2015→2017)

Largest single-wave increase in the dataset

31.5% → 35.0%

Teen persistent sadness inflection

The year the algorithm arrived

Suggestive timing alignment, not definitive causal identification. The feature-level resolution lets you see exactly what changed and when.

What we don't claim

The framework is honest about its boundaries. Here is what this paper does not prove:

N = 7 time points. R² values are high but confidence intervals are wide. This is proof-of-concept, not definitive.
Ecological design. Population trends, not individual-level tracking. The PISA individual-level dose-response (Paper 167, p = 0.007) addresses this.
Not standalone causal proof. We lead with external RCT evidence (Facebook’s own PNAS experiment, NBER deactivation studies) and use this as the specificity layer.
Every negative result is published. Electronic bullying: no signal. Global ecological correlation: null. We document failures alongside successes.
7/7 predictions confirmed. 5/5 kill conditions survived. The methodology puts itself at risk and has survived every test so far.