Social Media Feature Analysis
Papers 166/167 · Substrate: Epidemiological (CDC YRBS + PISA 2022) · Status: 12/14 PASS
The Circularity Break
The Void Framework's platform scoring (N=1,344) uses the framework's own rubric — a circularity concern. Papers 166/167 break this by replacing the rubric with verifiable facts about platform design: does the platform have an algorithmic feed? Autoplay? Opaque recommendations? These are binary/ordinal features anyone can check. The outcomes come from external health datasets (CDC YRBS, PISA 2022) that the framework had no role in collecting.
Same dimensional structure (Opacity, Reactivity, Coupling). Different operationalization. Independent data. No rubric. If the same structure predicts outcomes here, it's not circular.
The 13 Features
| Feature | Dimension | Type |
|---|---|---|
| Algorithmic feed | Opacity | Binary |
| Autoplay | Reactivity | Binary |
| Opaque recommendation | Opacity | Binary |
| Infinite scroll | Coupling | Binary |
| Like counts visible | Coupling | Binary |
| Follower counts visible | Coupling | Binary |
| Push notifications | Reactivity | Binary |
| Default public profile | Opacity | Binary |
| Ephemeral content | Opacity | Binary |
| Video-first format | Reactivity | Binary |
| Direct messaging (minors) | Coupling | Binary |
| Content creation tools | Coupling | Ordinal |
| Engagement gamification | Reactivity | Ordinal |
Result: Feature-weighted exposure R²=0.80 for persistent sadness (CDC YRBS, 7 waves). 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). Opacity features dominate (O avg R²=0.549 > R 0.493 > α 0.375). opaque_recommendation alone: R²=0.938 for female teen sadness.
Framework Prediction vs Result
The Void Framework predicted that opacity features would dominate harm outcomes. The data confirmed: Opacity > Reactivity > Coupling. The single most predictive feature — opaque_recommendation — is the purest opacity measure: the platform decides what you see and doesn't tell you why. R²=0.938 from one binary feature.
Litigation Relevance
This methodology is Daubert-qualified: verifiable features (any expert can check), large sample (613K students), cross-national replication (80 countries), external outcomes (CDC/PISA), and specific causal mechanism (opacity → explaining-away penalty → drift → harm). 14 predictions tested, 12 confirmed, 12/12 kill conditions survived.
The $6B+ social media litigation wave needs exactly this: a methodology that connects specific, checkable design choices to specific, measurable health outcomes, with a causal mechanism that doesn't require proving intent.
Caveats: Ecological inference (country-level exposure → individual outcomes) remains a limitation, though the PISA bridge (individual social media use → outcomes) provides within-country validation. 2023 pullback is attributed to COVID amplifier effect (LOO without 2021 R²=0.926). Monte Carlo robustness: 98.2% of 10K feature perturbations maintain R²>0.7.