Physical activity as an input — dose, fitness, and school outcomes

Verdict

This is the same null as motor competence & transfer, approached from the input side, and it does not become a different finding by being asked a different way. That file asks whether teaching movement spills over into cognition. This one asks the founder's version: how much activity, of what intensity, delivered how, and does fitness itself work as an input to school outcomes. Every one of those refinements has now been tested, and none of them rescues the claim.

What this topic adds is the detail the transfer verdict does not carry. The dose question has an answer and it is not "more." The fitted dose-response across randomized trials is an inverted U peaking at moderate intensity, topping out at g = 0.14 on laboratory executive function — and the longer the trial, the smaller the effect, with the three-year A+PAAC trial flat and the six-to-ten-week programmes carrying the entire positive signal. Acute and chronic separate cleanly and both fail for schools: single bouts move preadolescent executive function d = 0.02, and only 2 of 15 acute studies ever measured achievement at all. And the fitness-as-input model breaks at the join — interventions do raise cardiorespiratory fitness (SMD 0.24), but the fitness gain does not predict the achievement gain (β = −0.04, 95% CI −1.52 to 1.45).

The verdict is no-effect on cognition, achievement and g. It is emphatically not a verdict against physical activity, which has health, enjoyment and behavioural justifications this database does not evaluate — and which, per A+PAAC, costs nothing academically to provide.

What the evidence shows

Source	Design	Grade	Key effect
Wassenaar 2021 (Fit to Study)	preregistered cluster RCT, 104 schools / 18,261 pupils	A	Null on maths; EF SMD 0.017; null on fitness itself
Ciria 2023	umbrella review of 24 RCT metas	A	d = 0.22 → 0.13 (moderators) → 0.05 bias-corrected, CI crosses zero
Donnelly 2017 (A+PAAC)	3-year cluster RCT, 17 schools, WIAT-III	B	Null on maths, reading, spelling across 3 years; 55 min/wk MVPA delivered
Resaland 2016 (ASK)	cluster RCT, 57 schools / 1,129 children, national tests	B	Null, standardized diff 0.01–0.06 (p > 0.358); subgroup 0.62 in lowest tertile
Tarp 2016 (LCoMotion)	cluster RCT, 14 schools / 632 adolescents	B	Null on flanker EF and maths; control group better on interference RT (5 ms)
Solberg 2021 (School in Motion)	cluster RCT, 30 schools, national tests	B	Positive: numeracy d = 0.12 / 0.23, reading d = 0.06 / 0.18; attenuates under ITT
Bartoš 2025	study-level meta-meta, 2,239 estimates, RoBMA	B	EF SMD 0.012 (CrI 0.000–0.147); moderate-to-strong evidence AGAINST any effect in healthy youth
Hillman 2014 (FITKids)	RCT, 221 children, 9-month afterschool	B	Inhibition d = 0.27 (CI lower bound 0.00), flexibility d = 0.35 — lab tasks, no achievement outcome
Tuvey 2019	meta-regression, 17 trials	C	Fitness ↑ SMD 0.24; fitness change → achievement change β = −0.04 (CI −1.52 to 1.45)
Singh 2025 (BJSM umbrella)	umbrella, 133 reviews / 2,724 RCTs	C	EF SMD 0.24 uncorrected, "larger in children" — same data, 0.012 corrected
Vasilopoulos 2023	3-level meta, 92 RCTs / 25,334 children	C	No overall benefit on academic outcomes; on-task g = 1.04 at very low certainty
Li 2025	Bayesian dose-response NMA	C	EF g = 0.14; aerobic 0.22; inverted-U dose curve — more is not better
Chang 2012	acute meta, 79 studies	C	g = 0.097 overall; identical during / immediately after / delayed
Dixon 2025	acute meta, preadolescents	C	EF accuracy d = 0.02, RT d = −0.02; only 2 of 15 studies measured achievement
Álvarez-Bueno 2017	meta, 26 studies / 10,205 children	C	Maths 0.21, reading 0.13, language 0.16 (CI crosses zero), on-task 0.77
Singh 2019 (BJSM panel)	systematic review + expert panel	C	11 of 58 studies (19%) high quality; benefits in 48% / 60% of analyses — "inconclusive"
Donnelly 2016 (ACSM)	ACSM position stand	C	Achievement rated evidence category C, "neutral effect" by the field's own body
Marques 2018	systematic review, 51 articles (41 cross-sectional)	D	All 22 fitness studies positive; the 11 accelerometer studies "inconsistent"
Sardinha 2016	3-year cohort, 1,286 students, FITNESSGRAM	D	Fit-fit vs unfit-unfit OR 3.49 (Portuguese), 2.41 (English) — nothing in maths
Kyan 2018	within-person hybrid model, 567 students	C	Fitness→GPA boys only; girls null on both within- and between-person components

The trials get bigger and longer, and the effect gets smaller. Fit to Study (104 schools) and A+PAAC (three years) are the two most demanding tests ever run, and both are flat. ASK adds 57 clusters and independently scored national tests, and is flat. LCoMotion's one significant cognitive result favours the control group. Against that, School in Motion is the honest counter-example — but it is 10 clusters per arm, retrospectively registered, its effects run d = 0.06–0.23, and they attenuate under intention-to-treat with imputation.

The single most decisive number came out in 2025, and it came from re-analysing the pro-exercise case. Singh et al.'s BJSM umbrella pooled 133 reviews and 2,724 RCTs to claim "strong evidence ... across all populations," with the specific assertion that gains are larger in children and adolescents. Bartoš et al. rebuilt the same 2,239 estimates at study level and applied selection-model bias adjustment instead of trim-and-fill: executive function falls to SMD 0.012 (95% CrI 0.000–0.147), and in healthy children and adolescents specifically the analysis reports moderate-to-strong Bayesian evidence against an effect of any exercise category. That is evidence for the null in the exact population schools care about, not merely absence of evidence.

The observational fitness→achievement literature is large, consistent, and shaped exactly like a confound. Marques 2018 is the tell: all 22 cardiorespiratory-fitness studies support a benefit, 10 of 16 self-reported-activity studies do, and the 11 studies using accelerometers are "inconsistent." The exposure that correlates most reliably is the most heritable and least behavioural; the exposure that is actually manipulable and objectively measured shows nothing. Sardinha's cohort supplies the magnitude gap — persistently fit adolescents carry OR 3.49 for high achievement in their mother tongue — and the subject gap: the effect appears in language, not mathematics, which is the one subject every intervention meta claims responds.

Fitness does not function as an input. Tuvey's meta-regression is the cleanest test: across 17 controlled trials, physical activity reliably raised fitness (SMD 0.24) while moving achievement 0.11 (CI −0.16 to 0.38) and executive function 0.02 (CI −0.09 to 0.13) — and randomized changes in fitness did not predict changes in achievement (β = −0.04) or executive function (β = 0.26, CI crossing zero). The one trial that did find fitness mediation, School in Motion, found it only in the arm with the smaller academic effect and not at all in the arm with the larger one. A mediator that explains the weaker result and none of the stronger one is not a mechanism.

Acute effects are real, tiny, and not an achievement lever. Chang's lifespan meta puts the acute bout at g = 0.097, indistinguishable during exercise, immediately after, and after a delay. In the population that actually matters, Dixon's preadolescent meta finds d = 0.02 for accuracy and −0.02 for response time, with publication bias detected in a corpus of 15. And after five decades, only two of those studies ever measured academic achievement.

Hereditarian-lens assessment

Risk: low — the verdict rests on cluster-randomized trials with independently scored national and state tests, where genes cannot differ between arms. The heredity argument here is not a caveat on the verdict; it is the explanation for the literature the verdict overturns.

Cardiorespiratory fitness is 55–72% heritable with no detectable shared-environment component (Schutte 2016), physical activity 29–37%, and school achievement roughly 60% by adolescence. Fit and unfit children are therefore not two randomly assigned groups differing in an environmental input; they differ in genotype, body composition, family circumstance and conscientiousness, all of which independently predict grades. This is the textbook setup for passive gene–environment correlation, and it predicts precisely what the data show: a strong, consistent, cross-sectionally-largest association that collapses to zero the moment the exposure is randomized. OR 3.49 in a cohort versus d ≈ 0.00–0.23 in trials is not attenuation; it is a different quantity.

The design that would settle it does not exist. No within-family, discordant-twin, sibling-fixed-effects or Mendelian-randomization study of childhood fitness or physical activity on academic achievement was located in this survey. The closest available is Kyan 2018, a within-person hybrid model that absorbs every time-invariant characteristic of the child including genotype — and when that adjustment is applied, the uniform association reported by cross-sectional reviews survives only in boys and only on teacher-assigned grades. The most-confounded claim in this literature has never been tested with the design that would resolve it, which is a finding in its own right and mirrors what motor competence & transfer records about the Stodden cascade.

Boundaries & what critics say

• This does not duplicate motor competence & transfer — it converges with it. That file grades teaching movement skills and finds the spillover claims fail. This one grades activity as a dosed input and finds dose, intensity, duration, acute-vs-chronic and fitness all fail too. Where they differ: that topic carries genetic_confound_risk: medium because part of its verdict rests on longitudinal cascade evidence, whereas this verdict rests on randomized trials and so carries low. The two together close the loop from both directions.
• The exchange over Ciria 2023 is live and belongs on the record. Twenty-one leading investigators published a Nature Human Behaviour rebuttal (Dupuy et al. 2024) arguing the umbrella review omitted moderators for exercise type, intensity and dose, mishandled baselines, and might be a statistical artifact. Ciria et al.'s reply reports that adding every requested moderator left the bias-corrected near-null unchanged. We could not read either paper — neither deposits an abstract and both sit behind Nature's authentication gateway — so both are filed unresolved, and the characterization above comes from Sanabria 2026, which is written by an overlapping author group and should be read as such.
• The consensus is partly a citation artifact. Of 114 Scopus citations of Ciria 2023 as of April 2026, 31 (37%) miscite it — usually as reporting positive effects. The critical commentary has 30 citations to the reply's 8.
• The strongest positive results are the least trustworthy measures. On-task behaviour is the reliable finding in this literature (g = 1.04 in Vasilopoulos, 0.77 in Álvarez-Bueno) and it is a classroom-observation outcome rated very low certainty, four times any achievement effect. Children who have just moved settle better; that is a management fact, not a learning effect.
• The ASK low-achiever subgroup is the best remaining hope and it is fragile. Standardized difference 0.62 in the lowest baseline tertile is the most-cited positive result in the field, but it is one post-hoc interaction among many in a trial whose primary outcome was null, in the stratum most exposed to regression to the mean. The follow-up moderator paper makes it worse, not better: a three-way interaction at p = 0.044 in which high-performing girls trend negative. It has never been replicated in a trial designed to test it.
• FITKids is the strongest lab-based positive and it is not about school. Nine months, 221 children, inhibition d = 0.27 with a confidence-interval lower bound of exactly 0.00, against a wait-list control — the passive comparator whose adjustment shrank the pooled effect fourfold in Ciria 2023. It measured no achievement outcome at all.
• Fidelity failures are a real limitation on the mechanistic reading. Fit to Study delivered its protocol in ≥50% of lessons at only 33% of schools and did not shift objectively measured vigorous activity; LCoMotion also failed to move measured activity; A+PAAC delivered 55 of a targeted 100 min/week. These trials are strong evidence that scalable school programmes do not raise achievement, and weaker evidence about the underlying exposure–outcome relationship.
• Nothing here says activity is academically costly. The ACSM position stand and A+PAAC agree: adding physical activity did not depress achievement. Time spent moving is not time stolen from learning.

Practical guidance

• Budget physical activity as health, behaviour management and enjoyment — never as an achievement lever. Three years of daily active academic lessons moved achievement neither up nor down. That is the honest sales pitch: it is free, academically speaking.
• Do not pay a premium for dose, intensity or "optimal" protocols. The dose-response curve is an inverted U on a lab outcome, the largest apparent effects come from the shortest interventions (a small-study-bias signature), and the longest trial is flat.
• Do not set cardiorespiratory-fitness targets as a route to test scores. Trials raise fitness reliably and the fitness gain buys nothing academic. Fitness is worth having for its own reasons.
• Treat "fit kids do better in school" as a heritability fact until someone runs a within-family study. An OR of 3.5 in a cohort and d ≈ 0 in a trial describe two different things.
• Classroom activity breaks are defensible on on-task behaviour, and only that. State the claim at its actual strength — settled classrooms — and do not extend it to learning.
• Discount subgroup rescues. "It works for the low achievers" and "it works for boys" are both post-hoc findings from trials whose primary outcomes were null.

Open questions

• No genetically informative design has ever been applied to fitness or activity → achievement in children. A discordant-twin or sibling-fixed-effects analysis on any of the large fitness-testing datasets (FITNESSGRAM-linked administrative records would do it) is cheap, obvious, and missing.
• Whether the ASK low-achiever effect is real is testable and untested: a trial powered on that subgroup has never been run.
• Why School in Motion's non-aerobic "Don't Worry, Be Happy" arm outperformed its physically-active-learning arm on every outcome is unexplained, and it is the observation most damaging to every proposed aerobic mechanism.
• Whether physical activity affects attainment outcomes — attendance, engagement, persistence, graduation — separately from test scores is essentially unstudied, and is the outcome class where an effect would be most plausible given the on-task findings.
• Whether any of these programmes durably change activity habits (the health justification) is the same open question flagged in motor competence & transfer, and it remains unanswered.