Sleepmore in Seattle: Later school start times are associated with more sleep and better performance in high school students

94 sophomores measured in 2016 (pre) and 84 in 2017 (post), two Seattle public high schools; 81 and 76 respectively for non-school-day data. Watches were assigned to a subset of each class section "due to resource constraints", stratified to represent gender and underrepresented minorities.

US 10th-grade students at Roosevelt HS (31% economically disadvantaged, 7% ethnic minority) and Franklin HS (88% economically disadvantaged, 68% ethnic minority), Seattle

Pre/post around the Seattle School District's district-wide shift of secondary start times from 07:50 to 08:45 for 2016-17 (a 55-min delay). Wrist actigraphy (Actiwatch Spectrum Plus), >=5 school nights and >=3 non-school nights per student. Important design limits the abstract does not surface: the two waves are *different cohorts* of sophomores in the same 10th-grade biology sections, not the same students followed ("an independent sample of students taking the same science class was studied in each school"); and there is no untreated control school — the authors state plainly that "the nature of our pre-/post-study prevents us from having a control group". The internal control is the null result on non-school days measured in the same weeks, which is genuinely informative and is why this is graded above a bare pre/post. The academic outcome is second-semester grades in that same biology class, and the paper flags that these "represent absolute (not normalized) grades and could carry an implicit bias from teachers who could have been for or against the school time change" — the grade result therefore has neither a control group nor a blinded outcome measure. Grade lowered from B to C on this reading: METHODOLOGY grades "pre-post without control group" at D, and none of the grade-B designs (well-powered RCT, regression discontinuity, difference-in-differences with a clean shock, within-family) apply here.

Median daily sleep duration rose 34 min on school days (P = 0.0007, effect size 0.353) — from 6 h 50 min to 7 h 24 min — driven by a 44-min later median sleep offset (median wake 06:24 -> 07:08, P < 0.0001, effect size 0.194). Sleep onset showed "a trend toward a later sleep onset that was not statistically significant" (P = 0.0459 against a Bonferroni-corrected threshold of P < 0.017), so bedtime moved little but the delay was not fully absorbed; the authors note that a 55-min bell delay yielding only 34 min of sleep "suggests that after a year ... students may delay their bedtimes". Sleep parameters on non-school days did not differ between years, which is the internal falsification test. Social jet lag fell from a median of 1.60 to 1.25 h (P = 0.0118). Self-reported daytime sleepiness fell (median 7.0 -> 6.0, P = 0.0370). Median second-semester course grades rose from 77.5% to 82% (a 4.5% increase, P = 0.0261). Attendance and tardiness improved only at Franklin, the high-poverty school; Roosevelt showed no difference. Mood (BDI-II), chronotype, sex and school did NOT differ between years in any model. The authors themselves state it is "much harder to attribute causality for 4.5% higher grades on increased sleep".

Low for sleep duration: a district-wide bell-time change is assigned to every student regardless of chronotype or family, and the same-season non-school-day null rules out a generic measurement or cohort artefact. Higher for the grade result: the two cohorts are different students, and with no comparison school any district-level co-intervention, grading drift, or cohort composition change is fully confounded with the treatment.

Split the source in half. The SLEEP result is independently replicated: Widome 2020 (START, difference-in-differences with comparison schools, +41/+43 min), Roig 2026 (lottery-assigned shifts), Lo 2018 (Singapore, 45-min delay, ~10 min gain — same sign, one third the magnitude), and the Marx 2017 Cochrane/Campbell review (sleep positive in 6 of 6 studies). Magnitudes vary about threefold across settings, so the direction replicates and the effect size does not. The ACADEMIC result (4.5% higher median course grades) is NOT replicated: Hinrichs 2011 finds a flat null on ACT and on statewide tests in two further states, and Biller 2022 screened 21 studies for risk of bias and found no generalisable achievement effect. Do not let the replicated half vouch for the unreplicated half.