Preschool programs’ long-term effects, on trial.

No education policy argument leans harder on two small studies. When a minister says preschool pays for itself many times over, the claim traces back to a few dozen children in 1960s Michigan and 1970s North Carolina, followed for decades. The returns estimated from those cohorts — on the order of 7 to 10 percent a year — rival the stock market (Heckman et al., 2010).

Then there is the other result. Tennessee ran its state pre-K program through a genuine lottery, the strongest modern randomised test of scaled public preschool. The pre-K children left the program ahead. They lost the lead by the end of kindergarten. By sixth grade they scored below the control group, with more discipline referrals besides (Durkin et al., 2022).

It is tempting to pick one result and dismiss the other. That would be a mistake: both come from strong designs, honestly analysed. This article lays out the whole record: the famous model programs, the fade-out pattern, the national middle case of Head Start, and the sobering Tennessee trial. Then it asks what separates early investments that compound from early investments that evaporate.

Two experiments from another era

The HighScope Perry Preschool Project randomised 123 children in Ypsilanti, Michigan, between 1962 and 1967. All were Black, from low-income families, with low measured IQ at entry. The program itself was two school years of half-day preschool, run by certified teachers with small groups, plus a weekly home visit. The control group got nothing — in 1962, nothing was the normal alternative (Schweinhart et al., 2005).

The Carolina Abecedarian Project, begun in 1972, went further. It enrolled 111 infants from low-income families and provided full-day, year-round educational childcare from a few months old until age five — an order of magnitude more contact time than Perry (Campbell et al., 2002).

Both were tiny, expensive and obsessively documented. That last property is why they matter: the researchers kept following the children for decades. Perry’s cohort was interviewed at 19, 27 and 40. Abecedarian’s was assessed at 12, 15, 21 and beyond. Almost nothing else in education has an evidence trail that long.

The fade-out that fooled everyone

The early Perry results made headlines for the wrong reason. IQ scores jumped — a gap of roughly 11 points at the end of the program. Then the gap shrank every year, and by around age 10 it was essentially gone (Schweinhart et al., 2005). Critics reasonably called the program a failure. The gains it was designed to produce had washed out.

Fade-out turned out to be the rule, not a Perry quirk. Across roughly 84 program evaluations since the 1960s, the average effect at the end of a preschool program is about 0.35 standard deviations — and it shrinks steadily in the years that follow (Duncan & Magnuson, 2013). On achievement tests, the treated and untreated children converge. This is one of the most replicated patterns in education research.

If the story had ended there, preschool would be remembered as an expensive way to buy a temporary test bump. The story did not end there, because the Perry and Abecedarian teams kept counting — and switched from measuring test scores to measuring lives.

The adult ledger

By age 40, the Perry treatment group looked different where it counts. Roughly 77% had finished high school, against about 60% of the controls. Around 76% were employed, against about 62%. The arrest record tilted hardest: about 36% of the program group had been arrested five or more times, against roughly 55% of the controls (Schweinhart et al., 2005). Earnings ran higher too.

Abecedarian’s ledger, read at age 21, pointed the same way with different line items. About 36% of the treated group had enrolled in a four-year college, against roughly 14% of controls. Reading and maths advantages were still visible in early adulthood, and teenage parenthood ran lower (Campbell et al., 2002). Unlike Perry, some of Abecedarian’s cognitive advantage never fully faded — plausibly because it started in infancy.

Heckman and colleagues then put Perry through a proper cost-benefit audit, correcting earlier, looser claims. Counting earnings, crime, welfare and schooling costs, they estimated an annual social return of roughly 7 to 10 percent — benefit-cost ratios on the order of 7-to-1 to 12-to-1 under plausible assumptions (Heckman et al., 2010). Most of the return came from crime reduction and earnings, not test scores. The skills that persisted — self-control, persistence, social behaviour — were precisely the ones the achievement tests stopped seeing.

Head Start, the national middle case

Perry and Abecedarian were boutique experiments. Head Start is a nation-scale program that has served millions of low-income American children since 1965. It sits between the model programs and a modern state rollout — cheaper and more variable than Perry, older and better studied than any state pre-K. Its evidence splits cleanly along the same line as Perry’s.

On test scores, the verdict is sobering. The Head Start Impact Study randomised access for roughly 4,700 children — the program’s only national experiment. Offers of a Head Start place produced modest gains in the preschool year, mostly in literacy. By the end of third grade, the treatment and control groups were essentially indistinguishable on achievement (Puma et al., 2012). Fade-out, again, at national scale.

On life outcomes, the verdict flips. Deming compared siblings where one attended Head Start and the other did not, following them into adulthood. Test gains faded by adolescence, exactly on script. Yet by young adulthood the Head Start siblings scored about 0.23 standard deviations higher on a composite of graduation, college attendance, health and staying out of trouble. That is roughly 80% of the benefit of the model programs, at a fraction of the cost (Deming, 2009).

Hold the two Head Start results together and a rule emerges: the achievement test at age 8 is a poor forecaster of the life ledger at 25. Whatever the lasting asset is, third-grade scores do not price it.

Tennessee, the modern trial

Tennessee’s Voluntary Pre-K program is what scaled public preschool actually looks like: a full school-day program, thousands of classrooms, ordinary district staffing, real budgets. Oversubscribed sites ran admission lotteries, and Lipsey, Farran and Durkin followed about a thousand randomised children — the closest thing to a Perry-quality design ever run on a modern state program (Lipsey, Farran & Durkin, 2018).

The children left pre-K clearly ahead — roughly a third of a standard deviation on early achievement. The lead did not survive contact with school. By the end of kindergarten the control children had caught up. By second and third grade the lines had crossed on several measures, with the pre-K group slightly behind (Lipsey, Farran & Durkin, 2018).

The sixth-grade follow-up hardened the reversal. The randomised pre-K group scored lower on state achievement tests and logged more discipline referrals and more special-education placements than the controls (Durkin et al., 2022). The effects are small. They are also negative, in the strongest design available, at the scale ministers actually buy. No honest reading of the early-childhood literature gets to skip this study.

Both are real: reconciling the record

So the two headline results stand: 7–10% annual returns from 1960s Michigan, and a negative sign from 2010s Tennessee. Four differences carry most of the reconciliation.

The counterfactual moved. Perry’s control children stayed home in an era with almost no alternatives. A modern control group attends Head Start, subsidised childcare or an informal arrangement. Preschool evaluations increasingly measure “this program versus other care”, not “preschool versus nothing” — one reason recent programs show smaller initial effects than older ones (Duncan & Magnuson, 2013).

Quality and dosage differ by an order of magnitude. Perry ran tiny groups with certified teachers and weekly home visits; Abecedarian ran five years of full-day enrichment from infancy (Campbell et al., 2002). A scaled state program spends far less per child and varies enormously classroom to classroom. Averaging over that variation can average away the payoff.

What follows the program differs. Fade-out is fastest when the receiving school re-teaches what pre-K children already know, at one pace, from the same starting line. The Tennessee team themselves point in this direction: a head start only compounds if primary school builds on it (Lipsey, Farran & Durkin, 2018).

The outcomes measured differ. Perry looks transformative on lives and unremarkable on tests; Tennessee has so far been judged mostly on tests. That is not a defence of Tennessee — discipline referrals moved the wrong way too (Durkin et al., 2022) — but it cautions against reading any program’s obituary at grade 3 (Deming, 2009).

What buys a payoff

Strip the record to what correlates with lasting benefit and a short list survives. Intensity and duration: Abecedarian’s five full-time years left marks that Perry’s two half-days-plus-visits years did not, including cognitive gains that never fully faded (Campbell et al., 2002). Skilled adults and small groups, consistently. Content aimed at how children learn at that age — structured play, language-rich interaction, self-regulation practice — rather than a compressed first grade.

Targeting matters too. Every large payoff on record comes from children with the least at home; benefits shrink as the alternative improves (Deming, 2009). And the handoff decides retention: gains persist where the next classroom starts from each child’s actual level instead of the grade’s assumed one (Lipsey, Farran & Durkin, 2018).

None of this is exotic. It is expensive, per child, and it resists dilution — which is exactly why the model programs and the scaled ones keep producing different answers.

What the evidence doesn’t show

The early-childhood literature is long on conviction in both directions. The record itself supports neither triumph nor dismissal without conditions.

• Two cohorts carry the famous returns. Perry and Abecedarian together enrolled 234 children, in specific communities, half a century ago. The return estimates are careful, but they are estimates from tiny, era-bound samples (Heckman et al., 2010).
• Fade-out on tests is near-universal. No one should promise lasting achievement gaps from one pre-K year; the pooled record shows initial effects shrinking steadily after exit (Duncan & Magnuson, 2013).
• Sleeper effects are not guaranteed. Head Start’s adult payoffs emerged from quasi-experimental designs, not the national RCT — and Tennessee’s RCT shows follow-up can also reveal harm, not hidden benefit (Durkin et al., 2022).
• The mechanism is inferred, not proven. Self-regulation and related skills are the leading explanation for persistence, but no trial has randomised the mechanism itself (Deming, 2009).
• Universal programs are thinner evidence. The strongest payoffs come from targeted programs for the least-resourced children; extrapolating those returns to universal rollouts outruns the data (Duncan & Magnuson, 2013).
• Tennessee is one state’s implementation. The reversal indicts that program as run, and warns about scale — it does not prove every scaled pre-K fails (Lipsey, Farran & Durkin, 2018).

Where the evidence stops

1. 1Two cohorts carry the famous returns
2. 2Fade-out on tests is near-universal
3. 3Sleeper effects are not guaranteed
4. 4The mechanism is inferred, not proven
5. 5Universal programs are thinner evidence
6. 6Tennessee is one state’s implementation

© 2026 FUTURE PROOF™

Reading the evidence like a ministry

For a school system deciding what to fund, the record reads as five working rules.

Buy quality before coverage. The programs that paid off were intense, skilled and small-group; the one that backfired was scaled and variable. Expanding seats faster than quality is the one move the whole record argues against (Duncan & Magnuson, 2013).

Target first. The children with the least at home gain the most, in every strong study — Perry, Abecedarian and the Head Start sibling evidence agree on this even where they agree on little else (Deming, 2009). Universal ambitions should grow outward from the children the evidence is actually about.

Design the handoff, not just the program. Fade-out lives in the transition. A primary school that measures where each child actually is, and teaches from there, is the difference between compounding a head start and re-teaching it away (Lipsey, Farran & Durkin, 2018).

Judge programs on more than grade-3 tests — but do judge them. Adult ledgers redeemed Perry and Head Start after their test effects died (Schweinhart et al., 2005). Tennessee shows the opposite ending exists too (Durkin et al., 2022). Track behaviour, self-regulation and progression alongside achievement, and keep tracking after exit.

Embed the lottery. Every oversubscribed program is a free experiment. Tennessee’s willingness to randomise is why anyone knows the truth about it; that honesty should be the norm, not the scandal (Lipsey, Farran & Durkin, 2018). Fund the follow-up for decades — the entire Perry dividend was invisible at year five (Heckman et al., 2010).