Do later school start times increase adolescent sleep, and does that raise achievement?

Verdict

This is the cleanest causal evidence base in the archive, and the verdict is strong-support — for sleep, not for test scores. Almost every other topic here has to fight through selection: parents choose the programme, districts choose the curriculum, motivated students choose the practice schedule. A bell time is different. The district imposes it on every adolescent in the building simultaneously, regardless of chronotype, ability, or family. And the proximal outcome is measured by a wrist accelerometer rather than a questionnaire. Assignment is exogenous and the outcome is objective — the two things this archive almost never gets at once.

So do not deflate the sleep result. A 50-65 minute delay produces 41-43 extra actigraphy-measured minutes of school-night sleep, and it is still there two years later. Bedtime does not drift to absorb it.

The achievement claim is a different claim and it does not travel with the first one. The best-identified achievement estimates are 0.08 SD (math) and 0.06 SD (reading) per hour, a credible statewide ACT analysis returns a flat null, and the only systematic review that screened 21 studies for risk of bias found no generalisable effect under any grouping. That is a small, contested effect, not a large one, and the literature's own most careful reviewers describe the finding as achievement is likely maintained — a null-with-no-harm, not a lever.

What the evidence shows

The falsification tests are what make the sleep result trustworthy. Dunster's Seattle students showed no sleep difference on non-school days in the same weeks they gained 34 minutes on school days. Roig's lottery-assigned Argentine students showed the same pattern: the sleep gap between morning-shift and afternoon-shift boys disappears on nights before non-school days. In both cases the effect is present exactly where the constraint binds and absent exactly where it does not. That is not what selection looks like.

The achievement literature disagrees with itself, and the disagreement is not resolvable by picking a favourite. Edwards and Heissel & Norris are the two strongest identifications — a within-student fixed-effect design on bus-driven schedule changes, and an instrument built from a time-zone boundary — and both find real but small positive effects, 0.05-0.08 SD. Hinrichs has arguably the best outcome measure (individual ACT records for the universe of Twin Cities test-takers across a decade, replicated on Kansas and Virginia panels) and finds nothing. A defensible read is that the true effect is somewhere in 0.00-0.08 SD, which is a real but marginal number — comparable to a few weeks of schooling, not to the transformation the advocacy literature promises.

The behavioural and attendance effects are more robust than the achievement effects. James 2023 found 14% fewer behaviour referrals and three fewer late arrivals; Roig found roughly a 30-percentage-point drop in boys' aggressive/delinquent behaviour under random assignment; Sadikova found depression symptoms down 15.6% of an SD, with the effect concentrated entirely in the fatigue items. Note the honesty of that last decomposition: low mood, hopelessness, and worry did not move. This is a fatigue intervention that looks like a mood intervention on an aggregate scale.

Sex heterogeneity is real and under-discussed. Groen & Pabilonia found later starts gave girls more sleep and better reading, and gave boys neither. Roig found the mirror image — boys gained an hour, girls offset the early schedule with earlier bedtimes and naps and gained nothing. Both are consistent with a single underlying rule: the achievement effect appears only where the sleep effect appears, which is what a genuine mediated mechanism should look like, and which also means population-average effects understate the effect on whoever is actually constrained.

Hereditarian-lens assessment

Risk: low, and it is worth being precise about why, because this topic is the exception that proves the archive's rule.

Sleep duration is 69% heritable in adolescence (Kocevska 2021) — higher than at any other point in the lifespan, and higher than most of the outcomes it is claimed to cause. Chronotype is heritable too. So the observational sleep literature is exactly as compromised as this archive's premise predicts: families that produce short-sleeping teenagers also transmit alleles. The relevant correlation, in a meta of 15,000+ children, is r = 0.069 between sleep duration and school performance — about half a percent of variance, before any correction for heritable ability. That is the confounded upper bound, and the causal effect can only be smaller.

What breaks the confound is that a district-wide bell-time change is not chosen by anyone whose genotype matters. The Seattle school board did not consult students' chronotypes. Wake County's schedule shifts were driven by bus logistics. The Florida time-zone boundary was drawn for reasons entirely unrelated to child ability. Argentine middle schools assign the morning or afternoon shift by lottery. In every one of these designs, the treated and untreated groups have the same expected distribution of sleep-relevant alleles, and the treatment is a constraint imposed on the genotype rather than an expression of it. Edwards 2012 goes further and compares the same child to himself across a bell-time change, which nets out every stable genetic and familial influence by construction.

Where residual confounding still lives:

• Districts choose to change. The districts in START, Seattle, and Denver decided to move their bell times; districts that decide this are not random, and they may be doing other things at the same time. The comparison-school structure in START handles this; Dunster's Seattle design does not.
• Families move. Two- and three-year follow-ups allow selective relocation into or out of policy-change schools. None of the studies reports a serious test of differential attrition.
• Grades are assigned by teachers who know the policy changed. Every positive achievement finding on GPA or course grades (Dunster, James, Bastian & Fuller) is on a teacher-assigned measure. The independent standardized measures — ACT (Hinrichs), Florida state tests (Heissel & Norris), NC end-of-grade tests (Edwards) — give the smaller and more contested picture. This is the standard 2× measure-type inflation showing up exactly where the theory predicts.
• Astill 2012's null on intelligence is a useful boundary marker (recorded as an excluded source): in 86 studies of 35,936 children, sleep duration correlates r≈.08 with cognitive performance and school performance but shows no association with intelligence. Whatever sleep does, it operates on state-dependent performance, not on g. Nothing here should be read as a claim about ability.

Boundaries & what critics say

• No RCT exists and probably never will. Teensleep set out to randomize 100 UK schools to a 10:00 start and recruited two. Schools will not cede scheduling control, and the ones that would volunteer are self-selected outliers. This is why the Cochrane review rated the entire body of evidence very low certainty despite the consistent direction — and why the Argentine lottery, an accident of a double-shift system, is the most valuable single design in the topic.
• The strong circadian argument is overstated. The advocacy framing says adolescent bedtime is biologically fixed, so only wake time is manipulable. Campbell 2023 falsifies the strong version: in a within-subject PSG experiment, adolescents given more time in bed by advancing bedtime slept 68 more minutes going from 7 to 8.5 h and 57 more going from 8.5 to 10 h, with no change in this ability between ages 10 and 21. Bedtime is not immovable under an enforced schedule. What is true — and what the field studies actually show — is that under unenforced real-world conditions adolescents do not advance their own bedtimes, so the wake-time lever is the one that works in practice.
• Somebody pays. Districts almost always buy a later high school start by moving elementary earlier. Meltzer 2021 measured both ends: MS +29 and HS +45 minutes, but elementary students moved 60 minutes earlier lost 11 minutes. The trade is favourable — the loss is a quarter of the gain, and elementary self-reported sufficiency and sleepiness did not change — but it is a trade, and studies that report only the secondary-school side are reporting half a result.
• Carrell 2011's Air Force Academy result is the most-cited and least-applicable evidence. It is a genuinely randomized design showing a 50-minute delay worth a 1 SD improvement in teacher quality — on 18-year-old military cadets in a residential institution with enforced lights-out. Recorded as off-scope.
• McKeever & Clark's graduation-rate result should not be cited. Twenty-nine self-selected schools, pre-post, no control group, during a period of nationally rising graduation rates. Recorded as design-inadequate.
• Effect sizes here are not the effect of sleep. They are the effect of a bell time, which changes sleep, meal timing, light exposure, first-period alertness, after-school work and sport schedules, and family logistics all at once. Sadikova 2024 is the only source that attempts to isolate the sleep channel by instrumenting for it, and its outcome is mood, not achievement.

Practical guidance

• Start secondary school at 8:30 or later. The proximal effect is real, large, immediate, cheap, and persists for at least two years. Nothing else in this archive costs a scheduling change and delivers 40 minutes of measured behaviour change per student per night.
• Justify it on sleep, attendance, behaviour, and mood — not on test scores. Those are the outcomes with the strongest designs behind them (14% fewer behaviour referrals, one fewer absence, ~30pp less delinquency under randomization, depression symptoms −0.16 SD). Staking the policy on achievement makes it vulnerable the moment somebody reads Hinrichs or Biller.
• If someone forecasts an achievement gain, budget 0.00-0.08 SD. Not zero, not transformative. The effect is plausibly concentrated in lower-achieving students (Edwards: 2× larger in the bottom third) and in whichever subgroup is actually sleep-constrained.
• Count the elementary cost explicitly when the delay is financed by flipping the bus tiers. An 11-minute loss for 6-year-olds against a 45-minute gain for 16-year-olds is a good trade; making it silently is not.
• Do not substitute a sleep-education programme for the bell change. Nothing in this evidence base supports telling adolescents to go to bed earlier as a policy: Campbell shows they can under an enforced schedule, and every field study shows they don't. The constraint is what moves, not the advice.
• Where geography allows, weight start time against sunrise, not the clock. Heissel & Norris identify the effect off hours of daylight before school, and Rodríguez Ferrante's meta-regression confirms the solar clock moderates the effect. The same 8:00 bell is a different intervention at the eastern and western edges of a time zone.

Open questions

• Confidence is capped at medium pending the adversarial pass, not by the evidence. Three independent grade-B sources agree on the sleep-duration effect, which would clear the bar for high; status: surveyed means the skeptic review that METHODOLOGY requires has not been run.
• Nobody has measured this on an independent standardized test after a district change with a proper comparison group. START/LEARN reports GPA; Hinrichs reports ACT but off older policy changes. The obvious study — actigraphy plus state test scores plus comparison schools in the same panel — does not exist.
• Attainment is essentially unmeasured. Graduation and college-going are the outcomes that matter most and the only evidence is a pre-post with no control group. Given how often this archive finds attainment effects that survive test-score fadeout, this is the biggest gap.
• Why do the sex asymmetries reverse across contexts? Groen finds girls responsive and boys not; Roig finds the reverse. If the rule really is "achievement follows sleep", the interesting question is what determines who is constrained.
• How much of the effect is sleep and how much is everything else a bell time moves? Only one source attempts to instrument for the sleep channel specifically, and it does not look at achievement.
• Does the effect fade? Uniquely in this archive, the available evidence points the other way: both Widome's sleep effect and James's GPA and absence effects were larger in year two. Whether that is real accumulation, implementation maturity, or differential attrition is untested.