Preschool at scale — Head Start, state pre-K, and what universal provision delivers

Verdict

This is the evidence that should actually govern a preschool decision, and it is uncomfortable for every side of the debate.

Scaled pre-K does not durably raise achievement. Three of the four best designs say so, and one says worse than that:

• Head Start (national RCT, n=4,667): small end-of-year gains in narrow language/literacy, gone by kindergarten for 4-year-olds, and by grade 3 essentially nothing survives multiplicity correction. Both arms remain ~0.5 SD below national norms.
• Tennessee VPK (lottery RCT, n=2,990): positive at end of pre-K, gone by kindergarten, negative by grade 6 (−0.128 ELA, −0.178 math, −0.132 science), with more special-education placements (11.7% vs 8.4%) and more disciplinary infractions.
• Quebec (universal expansion): persistent negative non-cognitive effects into school age, worse adult health and higher crime.
• Boston (lottery): a precisely estimated zero on achievement across grades 3-10 — ruling out effects above ~0.12 SD.

And yet the attainment story is real. Boston, on the same precise zero, shows +8.3pp on-time college enrolment and +6.0pp high-school graduation. Three independent designs (Boston lottery, the Head Start rollout, the Ludwig-Miller RD) find attainment gains without test-score gains. This is the same decoupling seen in Perry — and it is now established in a modern, at-scale, lottery-based programme, which is what makes it credible.

So: mixed, at high confidence. Not "preschool doesn't work" — but "preschool does not do the thing it is usually sold as doing."

What the evidence shows

Source	Design	Grade	Key finding
Durkin 2022 (TN-VPK)	lottery RCT, n=2,990	A	End of pre-K +0.24 → K +0.02 → grade 6 −0.13 to −0.18; IEPs 11.7% vs 8.4%
Gray-Lobe 2023 (Boston)	lottery	A	MCAS grades 3-10 +0.005/+0.029 (precise zero); on-time college +8.3pp; HS grad +6.0pp
Puma 2012 (HSIS grade 3)	national RCT	A	One favourable result (Reading +0.11) failing multiplicity; one unfavourable; both arms ~0.5 SD below norms
Kline & Walters 2016	HSIS decomposition	B	vs home care +0.37 SD; vs other centre ~ZERO
Pages 2020	replication	C	Deming's +0.23 → +0.17 in his cohorts, −0.15 in later ones, ~0 combined
Watts 2023 (NC Pre-K)	county-dose TWFE, n=1.2M	C	The dissent: ~+0.20 SD growing through grade 5 — but the weakest design, internally contradictory

The counterfactual is the ballgame. Kline & Walters is the most clarifying paper here: Head Start beats staying home (+0.37 SD) but is statistically indistinguishable from the other preschools it displaces. As alternative childcare has become near-universal, the measurable return to adding a programme has shrunk toward zero. Much of what looks like "fadeout at scale" is really this.

But substitution cannot explain everything. Boston's compliers came ~62% from other preschool (33% Head Start, 29% private) — which should have made its effects smaller — yet Boston is where the attainment gains appear. Tennessee's controls were 63% home care — which should have made its effects larger — and Tennessee went negative. The counterfactual story runs the wrong way here, and the divergence between Boston and Tennessee remains genuinely unexplained. Structural quality doesn't explain it either: Tennessee met 9 of 10 NIEER quality benchmarks.

On the "no scaled programme shows persistent gains" claim — that is false as stated, and the adversarial review caught it. North Carolina reports gains that grow through grade 5. But it is a population-level, county-funding-dose design with an IV-rescaled effect size, its two constituent papers contradict each other on special-ed and retention, and it is outvoted by two cleaner randomised designs. Record it as the live dissent, not the answer.

Head Start's long-run attainment claim has failed replication. Deming's +0.23 SD becomes −0.15 in later cohorts and ~zero combined. Anyone citing Deming without Pages is citing half the evidence.

Hereditarian-lens assessment

Risk: low — the verdict rests on lotteries and RCTs. The lens contributes three things:

• Head Start's effect is floor remediation, not ability gain: >1 SD at the bottom of the distribution, ~0.25 SD at the top. Schools remediate the same floor for everyone within a couple of years, which is why the mean effect vanishes.
• The attainment/ability split is architectural. Shared environment for ability collapses to ~0.10 by adulthood; shared environment for attainment stays much larger. The observed pattern (test scores revert, graduation doesn't) is exactly what that predicts — see fadeout & persistence.
• No IQ test was ever administered in HSIS. PPVT is the g-loaded proxy, and notably it showed the smallest effect of any measure (+0.09) while parent-reported outcomes showed the largest (+0.31). That ordering is diagnostic.

Practical guidance

• Audit your counterfactual first. If the families you serve already have adequate preschool, expect ~zero measurable academic return. If their alternative is genuinely home care, expect ~+0.37 SD at end of treatment — and expect most of it to fade.
• Do not promise achievement gains. The best randomised evidence does not support them, and Tennessee shows a scaled, structurally high-quality programme can end up behind.
• If you run one, aim it at trajectory — school readiness that keeps children out of special-education placement and retention, and the behavioural/attainment channel that Boston moved.
• Watch for harm at scale. Tennessee and Quebec are both large, well-identified, and negative. This is not a "worst case it does nothing" intervention.
• The health channel is real and undersold — Head Start's most robust short-run effect was dental care (+15-16pp), and the Ludwig-Miller RD finds substantial child-mortality reductions.

Open questions

• Why Boston and Tennessee diverge is the central unanswered question in early childhood. Counterfactual, structural quality, and era all fail to explain it.
• Whether Boston's attainment gains replicate anywhere else with lottery-grade identification.
• Whether North Carolina's growing gains survive a stronger design.
• Boston's college graduation effect (+5.2pp) is not significant — the durable-attainment claim rests on enrolment, which is weaker than usually reported.