Preschool: Much More Than You Wanted To Know
Read the original on Slate Star Codex →
Follows up on
↳ Preschool: I Was Wrong — Essay · Nov 2018
Summary
An exemplary deep-dive on whether preschool/Head Start works. Setup: good RCTs show test-score gains fade out, but old intensive programs (Perry, Abecedarian) showed better ADULT outcomes; modern Head Start's only good RCT (HSIS) confirms fade-out but is too young for life-outcome data, so we lean on quasi-experiments. Scott walks through them with real rigor: Deming 2009 (his favorite — sibling controls on the NLSY, 30 pre-study variables checked, only birth weight differs and he corrects for it; +0.23 SD on a young-adult-outcomes index, ~8% more HS graduation, concentrated in at-risk subgroups; rare robustness checks); GTC 2002 and Currie-Thomas (weaker, contradictory subgroup patterns — billed as replications of Deming but actually disproving his specifics); the Brookings update (yet another racial pattern); Ludwig & Miller's regression discontinuity at the 300-poorest-counties cutoff (~5% HS-completion jump, at the edge of detectability); Carneiro & Ginja (less obesity/depression, but no HS effect, buried). Plus Chicago Child-Parent Centers, Havnes & Mogstad (Norway), Quebec (worse — the daycare-quality-vs-parental-care model), Chetty's kindergarten fade-out-then-earnings, and a 22-study meta-analysis (+0.24 SD HS graduation, p<0.001). (IV) The honest synthesis: everything is summed as 'Head Start works!' though each study finds small effects for different subgroups/outcomes (Deming and GTC literally disprove each other's specifics) — explainable as noise near the significance threshold. The case for believing it (no publication bias since every dataset's been used; sanity-check patterns; can't easily bias out all harmful findings; the significant meta-analysis) vs against (the famous preregistration chart — positive trials dropping 60%->10% once cheating got harder — and all this research predates preregistration). Lands on Yudkowsky's evidence-permits-vs-forces and a candid 60/40 that preschool helps.
Why this score
Quality 80 · Excellent. Excellent (80). A rigorous, comprehensive, intellectually honest synthesis of a messy literature and a master-class in epistemic care (sibling-control vs RD methodology, the publication-bias/preregistration chart, evidence-permits-vs-forces, the candid 'I'm not confident in this table'). Among Scott's most valuable MMTYWTK deep-dives.
Claude’s paradigm shift 55 · Moderate. Moderate, high (55). A synthesis of existing studies, but the meta-level observations — that these 'replications' disprove each other's specifics, and the preregistration-cheating concern applied to the whole field — are sharp and fresh.
Real-world impact 3 · Moderate. Moderate (3). An influential, careful treatment of the preschool/Head-Start evidence (a major policy program) within the education-policy/EA discourse; its reach is intellectual.