A feed tuned to hold attention is the digital familiar poured out over a whole population — the engagement loop with every chaperone the old houses built struck off. So the framework treats a platform the way it treats any reference: ask whether its design manufactures what it shows you (the created reference) or points past itself, and read that off verifiable design features, not vibes — thirteen of them (how opaque the ranking is, how engagement-maximizing the loop is), scored without a rubric.
What actually holds (lead with this)
Three claims are defensible, and they are the ones to carry:
- A girls-specific dose-response. In within-country fixed-effects on PISA, each step up in social-media use costs girls ≈ −0.13 life-satisfaction, while boys sit at ≈ 0 (the gender interaction is the signal, not the main effect). This is the cleanest, least-confounded result in the file.
- An ordering: opacity over engagement. Among the thirteen features, the opacity of the system (how little the user can see of why they’re shown what they’re shown) ranks above raw engagement-maximization as the associated harm — which is exactly what the Fantasia Bound predicts (the penalty lives in what the channel hides, not in how busy it is).
- A modest, real out-of-sample gain. Leave-one-out cross-validation improves (RMSE 2.99 vs 3.65) — small, but earned on held-out data rather than asserted.
What does NOT hold (and why the page says so)
The headline numbers are descriptive only — NOT load-bearing
You will see R² = 0.80 (US, 7 CDC waves) and r = −0.648 quoted for this work. Do not lean on either. They are level/subgroup associations, and the honest reading deflates them:
- R² = 0.80 is “too good, therefore artifactual.” Published causal estimates for social media and well-being explain fractions of a percent of variance (Orben & Przybylski 2019); 0.80 is ~200× larger — the signature of aggregation inflation (80 countries × 7 waves on trending series) and overfitting (13 features against 7 observations guarantees a high R²). On first differences the effect is r ≈ 0.05, n.s., and the 13 engineered features add only ΔR² ≈ 0.10 over raw adoption alone.
- r = −0.648 is a Western-Europe subgroup. Globally, life-satisfaction R² ≈ 0.01 — essentially null. Across 16 outcomes, only 1 survives Bonferroni on ΔR².
- So: report R²=0.80 and r=−0.648 as descriptive, trend-/subgroup-confounded associations, and lead instead with the girls dose-response + the opacity ordering + the LOOCV gain. This is the book’s own rule — the kills are the credibility — turned on the framework’s own result. The over-good number is the dry hole; marking it is the map working.
And mark the deeper reason for caution, which this book names elsewhere: separating the platform made them unhappy from unhappy people use the platform more is the homophily–contagion confound — generically unidentified from correlation alone. The honest claims above survive because they lean on within-person/within-country structure and held-out prediction, not on a raw cross-sectional R².
Appears in: The Modern Mirror · The Egregore · The Homophily–Contagion Confound (why the causal read is hard) · The Fantasia Bound (opacity-over-engagement = the penalty ordering) · Divination (the chaperone-stripped loop).