Length of the school day and year, time-on-task, extended time, and summer

Verdict

Instructional time is not a production input with a constant return, and the literature only looks contradictory if you assume it is. Two questions get conflated. Does removing school time hurt? Yes, reliably, and it is one of the better-identified facts in this archive. Does adding school time help? Barely, expensively, with sharply diminishing returns, and only where the marginal hour lands somewhere that can use it.

The asymmetry is not a paradox. Removing time removes hours that were already being spent on the highest-value use available; adding time adds hours at the bottom of a school's priority list, into whatever slot is left, with whatever teacher is available at 4pm. Lavy's own concavity says so — the marginal hour in the two-to-three-hour range is worth 4.20 PISA points, the marginal hour past four hours only 2.48 — and Rivkin and Schiman say the rest of it: the return more than doubles between a disorderly and an orderly classroom, and schools in the bottom tail get "little or no benefit."

The headline number in this literature has already failed a replication and almost nobody has noticed. Lavy's 0.058 SD per weekly subject-hour is the most-cited estimate of the value of instructional time. Bietenbeck and Collins ran his exact specification on eleven further waves of PISA and TIMSS in the same 22 countries: they reproduce 0.058 exactly in PISA 2006 and get 0.014 to 0.037 everywhere else, tracing the gap to PISA 2006 having measured instruction time in categorical bands. Rivkin and Schiman, working independently with school-by-subject fixed effects, land at 0.017 to 0.025. The defensible central estimate is about 0.02-0.03 SD per weekly subject-hour, not 0.058, and certainly not the 0.15 that gets quoted when the within-pupil standard deviation is used as the denominator.

What the evidence shows

Taking time away

Adding time back

Summer

Hereditarian-lens assessment

Risk: low, and unusually cleanly so. Nearly every estimate above comes from a shock imposed on children by an institution: a district changing its calendar for budget reasons, a state mandating an extra hour in its lowest-scoring 100 schools, a government closing every school in the country, a military bureaucracy assigning an enlistment date from a birthdate and a parish. None of those decisions is made by anyone whose genotype is correlated with the outcome. Carlsson and colleagues go furthest and demonstrate conditional randomness directly, showing that days of schooling are uncorrelated with ninth-grade grades, parental education and paternal income once the assignment variables are conditioned on.

Two places where the confound has not been removed, and they are the two places this literature makes its most political claims:

• The socio-economic gradient in summer loss. Cooper's reading result — middle-class children gaining and disadvantaged children losing — is a between-family comparison of what happens when school stops constraining the environment. Families differ in books, supervision, travel and parental ability, and they differ in alleles, and a seasonal comparison cannot separate them. The strongest datasets now say the gradient is small anyway: race and SES together explain about 4% of the variance in summer loss.
• The COVID socio-economic gradient, for the same reason. That disadvantaged children lost more is well-established; that home environment caused the difference is not identified by any study in the meta-analysis.

Carlsson's crystallised/fluid split is the archive's premise showing up inside a single well-identified study. Ten extra school days move synonyms and technical comprehension; they move spatial and logic tests not at all, and if anything negatively relative to a non-school day. Schooling buys taught knowledge. It does not buy general ability. That is exactly the distinction Ritchie, Bates and Deary draw from the other direction, and it is why "a year of school raises IQ by 1-5 points" and "no pedagogical intervention durably raises g" are both true.

Boundaries & what critics say

• The famous number is scale-dependent, and three different numbers describe one coefficient. Lavy's estimate is 5.76 PISA points, which is 0.15 of the within-pupil SD (38.8), 0.07 of the between-pupil SD (84.4), and 0.058 of the 100-point international scale. Advocacy quotes the largest. Any claim about "the effect of an hour" that does not name its denominator is unusable.
• Snow days do not cost what absences cost, and this is the deepest result about time in the archive. Goodman shows the earlier snow-day literature was mis-identified because snowfall moves both closures and individual absences; separating them, each snow-induced absence costs 0.05 SD of maths while each closure day costs an amount statistically indistinguishable from zero, with effects larger than 0.01 SD/day ruled out. Synchronised time loss is absorbed by the teacher; staggered time loss is not. Time is not a quantity of hours, it is a quantity of coordinated hours.
• The COVID estimate is not a clean subtraction of instructional hours. Closure substituted remote schooling for in-person schooling and arrived with household disruption, illness and economic shock. It bounds what losing normal schooling costs; it does not price an hour.
• The Danish RCT's ranking is uncomfortable for everyone. Unstructured extra time (+0.15 SD) beat expert-programmed extra time (+0.04, n.s.), and the authors conclude "a general increase in instruction time is at least as efficient as an expert-developed, detailed teaching program." The arms could not be statistically separated, so this is a null result about programme design rather than a finding that structure hurts — but it should stop anyone claiming the extra hour only works if it is scripted.
• The summer-loss dispute is genuinely two disputes, and honest reporting has to keep them apart. Von Hippel and Hamrock demolish the measurement of the classic result: the Baltimore study used Thurstone-scaled fixed-form tests whose scale spreads with age, and estimated school-year learning within a test form while estimating summer learning across two different forms. On IRT scales, average gap growth from grade 1 to grade 8 is 7 percent and no gap doubles. But they do not deny summer loss: "the figures do consistently show that summer learning is slow for nearly all children, including children from advantaged groups." Atteberry and McEachin, on adaptive tests immune to the form artifact and with reconstructed district calendars, find mean loss at every grade and put it at a fifth to a third of the school-year gain. The mean loss replicated. The gap story did not — and Kuhfeld, Condron and Downey find the Black-White gap does the opposite of what the canon says, widening during school and narrowing over summer because White students lose more.
• A third mechanism in that dispute is still unquantified. Kuhfeld notes that students disengage from testing more in autumn than in spring, and that autumn tests are typically administered four to eight weeks into the year. Nobody has bounded how much apparent summer loss is test-taking effort. Nor has anyone bounded regression to the mean, which Kuhfeld shows is the single strongest predictor of summer change: prior-year growth alone explains 22-39% of the variance.
• Design quality predicts effect size here as reliably as anywhere in the archive. Lynch and colleagues report summer studies with no control group averaging 0.30 SD against 0.09 SD with one. Kidron and Lindsay explain Cooper's 0.26 SD summer-school result as an artifact of "mostly studies with less rigorous study design." The RAND lottery gives 0.08 SD, fading. The pattern is monotone.
• The one unambiguous win is targeted, not general. Chicago's double-dose algebra doubled instructional time in one subject for below-median freshmen and produced 0.18-0.24 SD on independent tests, +9.8pp graduation and +3.3pp BA attainment twelve years later — and then produced a null in the following cohort when cut-score adherence slipped and peer sorting worsened. Extra time works when it is aimed at a specific skill deficit in a specific group. It does not work as a longer day.

Practical guidance

• Protect the time you have before buying more. Every subtraction experiment finds a cost; most addition experiments do not find a benefit. A four-day week is a 3.6-hour weekly cut for 0.03-0.06 SD; individual absences cost about 0.008 SD each within-student.
• If you add time, add it to a subject and a group, not to the day. Chicago's double-dose algebra and Florida's literacy hour are the two positive results, and both are subject-specific and targeted. Massachusetts's 300 extra hours a year for everyone, sustained five years, bought nothing and cost engagement — more students reported being tired and fewer liked being at school.
• Budget 0.02-0.03 SD per weekly subject-hour, and less if your classrooms are disorderly. If a proposal forecasts more than that, ask which standard deviation it is dividing by.
• Fix the classroom before extending it. Rivkin and Schiman's interaction is the most policy-relevant number in this topic: a school in the bottom quartile of classroom order gets roughly nothing from extra hours. Extra time is a multiplier on whatever is already happening.
• Expect real summer forgetting and do not sell summer school as gap-closing. Plan for a fifth to a third of the year's gain to come back off, mostly in maths. But the achievement gap does not open in summer, so a summer programme justified on equity grounds is being justified on a claim that the best data contradict. Randomised summer programmes buy about 0.08-0.10 SD and it fades.
• Never justify anything here as raising ability. Schooling durably raises measured IQ by 1-5 points a year and it does so entirely through taught skills; Carlsson's fluid-intelligence nulls are the cleanest single demonstration.

Open questions

• Confidence is capped at medium by process, not evidence. Several independent grade-A/B sources agree on the core asymmetry, which would clear the bar for high; status: surveyed means the adversarial pass METHODOLOGY requires has not been run.
• Nobody has priced a marginal hour with random assignment at scale. The Danish trial (+0.15 SD for 3 h/week over 16 weeks) and the Dutch trial (nothing for 5 h/week over 3 months) disagree, and both are small and short. The obvious experiment — randomise schools to +2, +4 and +6 weekly hours and follow them two years — does not exist.
• How much apparent summer loss is test-taking effort? The mechanism is named by the field's own leading measurement expert and has never been bounded. Until it is, the 17-34% figures are upper bounds.
• Why did extended-time programmes fail where targeted double-dosing succeeded? The candidates — who gets the extra hour, what fills it, whether the teacher is fresh, whether the group is ability-matched — are all testable and none has been tested against the others.
• Almost nothing here reaches attainment or earnings. Pischke's German earnings null is precisely estimated and is essentially the only long-run evidence on school-year length. Given how often this archive finds attainment effects that outlive test-score fadeout, that is the largest gap.