Feature Types
Opacity (O)
Reactivity (R)
Coupling (α)
Paper 173· CC-BY 4.0· Social Media Features· DOI: 10.5281/zenodo.19455974

We Can Name the Exact Features That Hurt Teens

Not “social media is bad.” Which features. Which design choices. 13 binary flags scored from public changelogs and press releases. No expert rubric. Just checkable facts.

See the Features ↓ Read the Paper →
R² = 0.938
A single design feature — opaque recommendation — explains 93.8% of the variance in female teen persistent sadness across platforms.
Feature-weighted exposure across all 13 features: R² = 0.80. CDC YRBS data, 7 waves (2011–2023).

The 13 features

Every feature is a binary or ordinal fact about a platform’s design. Not a subjective rating — a verifiable design choice documented in changelogs, patents, or press releases.

FeatureTypeR² (female sadness)
opaque_recommendationOpacity
0.938
algorithmic_feedOpacity
0.720
autoplayReactivity
0.650
hidden_rankingOpacity
0.580
infinite_scrollReactivity
0.540
social_comparisonReactivity
0.510
opaque_moderationOpacity
0.490
push_notificationsReactivity
0.480
variable_rewardsReactivity
0.440
engagement_metricsCoupling
0.420
default_publicCoupling
0.380
ephemeral_contentCoupling
0.320
streak_mechanicsCoupling
0.290
Opacity features dominate: average R² = 0.549 (O) vs 0.493 (R) vs 0.375 (α). The features you can’t see are the ones that hurt most.

Two independent datasets

The same features, tested against two completely separate datasets from two different populations. Both confirm the pattern.

R² = 0.80

CDC YRBS (United States)

7 waves, 2011–2023. Feature-weighted exposure predicts persistent sadness in female teens. The dataset is public. The methodology is reproducible.

Youth Risk Behavior Survey. National sample.

613,744

PISA 2022 (80 countries)

Cross-national replication. r = −0.648 in Western Europe (p = 0.017), survives GDP control. Girls 5.6× more affected in 91% of countries (p < 0.000001).

Programme for International Student Assessment. OECD.

Bradford Hill causation criteria

The gold standard for establishing causation in epidemiology. 9 criteria. We satisfy 8.

SATISFIED

Strength

R² = 0.80 composite. opaque_recommendation alone: 0.938.

SATISFIED

Consistency

Replicated across CDC (U.S.) and PISA (80 countries).

EXPECTED MISSING

Specificity

Multi-feature exposure cannot be specific to one outcome. This is expected and predicted.

SATISFIED

Temporality

Feature rollouts precede outcome changes in time-series data.

SATISFIED

Biological gradient

More features = more harm. Dose-response confirmed.

SATISFIED

Plausibility

Opacity features hide mechanism. Reactivity features sustain engagement. Coupling features prevent exit.

SATISFIED

Coherence

Consistent with attention research, developmental psychology, and information theory.

SATISFIED

Experiment

Natural experiment: platforms that removed features show reduced harm.

SATISFIED

Analogy

Gambling machine design features show identical pattern: opacity + variable rewards + engagement lock-in.

This is Daubert-qualified methodology. Verifiable features. External health data. Replicable by any expert. No framework rubric anywhere in the pipeline.

Go deeper

The features are public. The data is public. The methodology is reproducible. Check it yourself.

📄
Read Paper 173
Full analysis on Zenodo. CC-BY 4.0. All 13 features, both datasets, complete methodology.
 🔒
The Ceiling (Paper 170)
Why opacity features dominate. The explaining-away penalty makes hidden mechanisms the most harmful design pattern.
 📊
All Evidence
Six non-circular confirmations. Social media is one of them. See the complete picture.
← All Papers All Evidence →