Head Start Impact Study: Final Report

4,667 randomized (2,559 age-3 cohort; 2,108 age-4 cohort)

Head Start-eligible children nationally.

The only large national RCT of Head Start — congressionally mandated, designed on the recommendation of the independent Advisory Committee on Head Start Research and Evaluation, sponsored by ACF (the agency that administers Head Start; the ACF project officer is on the author line) and executed under contract by Westat with Abt, AIR, Urban Institute and others. Two independently randomised age cohorts (entering at 3 and at 4), each followed to the end of 1st grade; the 3-year-old cohort has an extra measurement wave (end of the age-4 year), during which many treatment children were in a SECOND Head Start year. No IQ test was administered — PPVT (receptive vocabulary) is the most g-loaded proxy available. Roughly 10,000 impact comparisons were tested across two cohorts, four waves and many outcomes, so the report grades evidence in three tiers: STRONG (p<0.05 and survives Benjamini-Hochberg), MODERATE (p<0.05 only), SUGGESTIVE (p<0.10, may or may not survive). Nearly every effect size in the summary exhibits is reported at the p<=0.10 threshold, so several "impacts" are suggestive-only. No-shows ~15-20% and crossovers ~14-17%; both ITT and IV-adjusted participation impacts are reported, and everything below is ITT.

The definitive test of Head Start as actually delivered, and the cohort distinction is load-bearing. 4-YEAR-OLD COHORT: six small language/literacy gains at the end of the Head Start year (0.09-0.25 on direct assessments, 0.31 on parent report), never any math or pre-writing or promotion impact in any year, exactly zero cognitive impacts at the end of kindergarten, and at 1st grade only a suggestive PPVT +0.09; social-emotional zero in the Head Start year and kindergarten, then mixed and partly unfavourable at 1st grade. 3-YEAR-OLD COHORT: a broader end-of-year effect including math (Applied Problems +0.15) and PPVT +0.18, of which only two measures survive to the age-4 year and none survive to 1st grade except a suggestive Oral Comprehension +0.08; it is also the only cohort with social-emotional impacts, all of them parent-reported, and it carries the study's single unfavourable cognitive impact (teacher-rated math ability -0.19 at kindergarten). Across both cohorts, the durable findings are health-access and parenting, not learning: dental care +15pp (4yo) and +17pp (3yo) at the end of the programme year, gone by 1st grade, and health insurance coverage +4pp still present for the 4-year-old cohort at 1st grade. The counterfactual is an active alternative, not nothing, and both groups remain far below national norms throughout.

NOT nothing, and the report says so explicitly: Head Start was tested "against a mixture of alternative care settings rather than a condition of no services". 4-year-old controls (Exhibit 1, spring 2003): 39.7% parent care only, 35.3% other centre-based care, 13.8% crossed into Head Start anyway, 11.2% other care. 3-year-old controls: 38.4% parent care, 25.2% other centre-based, 17.8% Head Start crossover. Controls who were in non-parental care were in it 4-5 MORE hours per week than the Head Start group. In the 3-year-old cohort's second year the control group was given access to Head Start, so the age-4 contrast is badly contaminated: ~50% of controls in centre-based care were in Head Start by spring 2004.

RCT. The effect ordering is diagnostic: largest on parent-report, smallest on the most g-loaded measure.

No second national RCT of Head Start exists, so `unreplicated` refers to the design, not to the fadeout pattern, which is one of the most replicated regularities in the literature. The significant end-of-Head-Start-year cognitive effect sizes in this study run 0.09-0.31 (4yo cohort) and 0.10-0.35 (3yo cohort), centring near 0.15-0.20 — roughly a quarter to a third of the ~0.60 SD claimed for Perry/Abecedarian, which is the intensity gap that blocks extrapolation from the boutique trials. The ~0.14 SD "average" formerly recorded here is an external synthesis figure, not a number this report states.