Breakfast, school meals, and micronutrients — what feeding children actually buys

Verdict

"Nutrition and learning" is not one question, and treating it as one is how a real finding about deficient children became a false promise about everyone else's. Three questions have to be kept apart, and the evidence answers them differently:

1. Does correcting a documented deficiency improve cognition? Yes, moderately — and the gain is in intelligence-test and lab-task scores, not in school achievement.
2. Does supplementing an already-fed child improve anything? No. This is the closest thing to a clean null in the archive after motor-competence transfer.
3. Do school meal programs raise achievement? Yes, by 0.03–0.10 SD — but the designs that show it cannot attribute it to nutrition, and the mechanism evidence points elsewhere.

The verdict is mixed because the effect is real in one population and absent in another, which is exactly what mixed means. It would be no-effect if the question were only about well-fed children.

What the evidence shows

Source	Design	Grade	Key effect
Fiani 2025	meta of RCTs, non-anaemic	C	Intelligence d=0.46, memory d=0.53 — but effects absent when iron-deficient participants are excluded
Gutema 2023	meta of 13 iron RCTs, ages 6–12	C	Intelligence 0.46; anaemic at baseline 0.79 (0.41–1.16); school achievement 0.06 (−0.15 to 0.26), null
Gordon 2009	double-blind RCT, 184 mildly iodine-deficient NZ children	B	Overall cognition +0.19 SD (P=0.011); 2 of 4 WISC subtests moved; deficiency biochemically verified
Asgari 2026	meta, 35 RCTs, 19,343 children	C	Language 0.09 (GRADE high); memory 0.22 (GRADE very low); nothing else, including intelligence; 4/35 trials low risk of bias
Eilander 2010	meta of 20 micronutrient RCTs	B	Fluid IQ 0.14 (−0.02 to 0.29), ns; crystallized −0.03; academic 0.30 (0.01–0.58) on 4 trials
Warthon-Medina 2015	meta of 6 child zinc RCTs	C	Intelligence <0.001 (−0.12 to 0.13); EF 0.08; motor 0.11 — all null
Montgomery 2018 (DOLAB II)	preregistered replication RCT, 376 children, 84 schools	B	Failed to replicate: no consistent effect on reading, working memory, or behaviour
Richardson 2012 (DOLAB I)	RCT, 362 children, 74 schools	B	Null on reading in full ITT; positive only in the n=224 sub-20th-centile subgroup; parent-rated behaviour moved, teacher-rated did not
Seura 2025	meta of 24 observational studies	D	Breakfast skipping → OR 2.08 (1.82–2.37) for poor achievement; survives every covariate adjustment
Luan 2022	cluster RCT, 16 schools, 1,362 students, 2.5 yr	B	Reading 0.02 (SE .06); math −0.20 (SE .07), P=0.005; attendance 0.004
Imberman & Kugler 2014	DiD on rollout timing, large urban district	B	~+0.10 SD math and reading — same intervention contrast as Luan
Ni Mhurchu 2013	stepped-wedge cluster RCT, 14 NZ schools	B	Attendance OR 0.81 (0.59–1.11), null; achievement null; satiety +8.6 points
Frisvold 2015	RD + DiD on state SBP mandate thresholds, NAEP n=53,430	B	Math +0.077 (SE .032) DD / +0.091 (SE .044) RD; reading +0.054 / +0.122; first stage +33 pp
Anderson 2018	DiD on vendor-contract turnover, all CA schools × 5 yrs	B	Healthy vendor +0.036 SD; placebo (t−1) <0.01; ≤$222 per 0.1 SD; no obesity effect
Ruffini 2021	DiD on staggered CEP adoption	B	Breakfast take-up +38%, lunch +12%; math gains only where baseline eligibility was low
Kristjansson 2025 (Cochrane)	13 cluster-RCTs + 27 NRSIs, 91,885+ students	B	LMIC math SMD 0.14 (0.06–0.23, GRADE high); reading 0.02, null; enrolment +3.4 pp; attendance null; HIC evidence: one non-randomised study
Duan 2024	DiD on county rollout, 40.6M students covered	B	Math up, verbal null; authors attribute effect to health, attendance, study habits, expectations, peers
Adolphus 2016	systematic review, 45 intervention studies	C	Breakfast-vs-fasting helps attention/EF/memory same morning only, more in undernourished children; chronic programmes inconsistent
Adolphus 2017	methodological critique	D	Bespoke lab batteries, artificial settings, unblindable conditions, habitual-breakfast self-report — the field's own authors say the effect is not substantiated
Wolraich 1995	meta of 23 blinded placebo-controlled sugar challenges	B	All 14 outcome constructs have CIs including zero (range −0.14 to +0.30). Sugar does not affect behaviour or cognition

The single most informative number in this topic is a subgroup exclusion. Fiani 2025 pools iron supplementation trials in non-anaemic people and finds respectable effects — intelligence d=0.46, short-term memory d=0.53. Then it removes the iron-deficient participants, and the effects are absent. Gutema 2023 shows the same gradient from the other side: 0.46 SD on intelligence overall, 0.79 SD restricted to children anaemic at baseline. Iron is not a nootropic. It is a treatment for iron deficiency, and it behaves exactly like one.

Iodine is the cleanest demonstration that "deficient" is not a synonym for "poor country." The Dunedin trial randomised 184 New Zealand 10–13 year olds with a documented median urinary iodine of 63 µg/L, supplemented for 28 weeks, verified repletion biochemically, and measured cognition with WISC subtests rather than a bespoke battery. Overall cognition rose 0.19 SD. Two of four subtests moved and two did not — a pattern that would read as noise-mining in a trial without biochemical verification of the mechanism, and which is why this is a grade-B single trial rather than a settled finding.

Where no deficiency is documented, the supplementation literature is null and gets more null as it gets bigger. Eilander 2010 pooled 20 micronutrient RCTs and found fluid intelligence 0.14 SD with a confidence interval crossing zero. Asgari 2026 pooled 35 RCTs and 19,343 children — nearly an order of magnitude more data — and recovered a 0.09 SD language effect and nothing else, with only 4 of 35 trials at low risk of bias and the memory estimate rated GRADE very low. Zinc is 0.00 SD. The authors of the largest meta conclude their own data are insufficient to recommend supplementation.

The omega-3 story is a textbook subgroup-then-replication sequence and it ended badly for the claim. DOLAB I was null on reading in the full randomised sample and positive only among the 224 children below the 20th centile; parent-rated behaviour improved while teacher-rated behaviour did not, in a supposedly blinded trial. DOLAB II was run by the same authors, preregistered, with a published protocol, larger (376 children, 84 schools), and aimed squarely at the responsive subgroup. It found nothing. This is what a good-faith replication looks like and the correct inference is that the original subgroup effect was noise.

Breakfast is where the confounding gap is widest and most measurable. Pooled observational data put breakfast-skipping at roughly double the odds of poor achievement (OR 2.08), an association that survives every covariate adjustment the meta-analysts tried — which is the tell, not the reassurance. Randomised provision produces: reading 0.02 SD and math −0.20 SD over 2.5 years in Philadelphia; null attendance and null achievement in New Zealand, where the programme did successfully reduce hunger (+8.6 satiety points). The intervention hits its nutritional target and misses the educational one.

The one place the breakfast literature reliably shows an effect is the acute lab paradigm, and that is a different claim. Adolphus 2016 finds that a child who has fasted overnight does worse on attention, executive-function and memory tasks that same morning, with effects larger in undernourished children. That is a statement about being hungry right now. It does not license a claim about achievement, and Adolphus 2017 — the same group — catalogues why the field's measurement choices systematically inflate it: bespoke cognitive batteries administered under artificial conditions, unblindable treatments, and self-reported "habitual breakfast."

School meal programs do work, and the reason is not nutrition. Frisvold's regression discontinuity on state mandate thresholds finds 0.08–0.12 SD on NAEP — genuinely causal, with a 33-point first stage. Anderson's vendor-contract design finds 0.036 SD with a flat pre-trend and a cost of at most $222 per 0.1 SD, which makes it one of the most cost-effective interventions in the entire database. But look at what identifies each. Ruffini finds gains only where baseline free-meal eligibility was low — that is, where the policy transferred income to families who previously paid. Duan's mechanism decomposition names attendance, health, study habits, expectations, and peer composition. Cochrane finds enrolment rises 3.4 points while attendance does not budge and reading does not move. And Imberman & Kugler's +0.10 SD for Breakfast in the Classroom is contradicted by Luan's cluster RCT of the identical contrast, which found −0.20 SD in math. A program that feeds children is a good program. The design that proves it works does not, and cannot, show that the food is why.

Sugar is a real debunking. Wolraich pooled 23 blinded, placebo-controlled sugar challenges across 14 outcome constructs and every single confidence interval includes zero. Parental belief in the "sugar rush" is expectancy.

Hereditarian-lens assessment

Risk: low, because the verdict is built almost entirely from randomised trials, cluster RCTs, regression discontinuities and administrative-rollout natural experiments. Genes cannot differ across arms of a placebo-controlled iron trial or across a state-mandated threshold.

That low rating is a product of source selection, not a property of the field. The nutrition literature at large is among the most genetically confounded in this archive:

• Diet quality is a parental choice variable. Which children eat breakfast, eat vegetables, take supplements, and eat regular family dinners is determined by parental conscientiousness, organisation, income and IQ — all substantially heritable, all transmitted genetically to the child, and all direct predictors of achievement independent of anything on the plate. The parents who supply the balanced breakfast supply the alleles.
• Breakfast-skipping is a near-perfect proxy for household chaos. In adolescents it additionally indexes chronotype and parental supervision. An OR of 2.08 that barely moves under SES adjustment is what a stable common cause looks like; measured SES is a crude and noisy proxy for the heritable parental traits actually doing the work, so "adjusted for SES" removes very little of the bias. No study in that pooled literature uses a within-family, discordant-sibling, or genetically informed design.
• RCTs of a nutrient cut through this completely. Randomising a pill or a fortified food breaks the link between parental traits and the child's exposure. This is why the supplementation literature is trustworthy and the diet-quality literature is worthless for causal purposes — and why they disagree.
• RCTs of a program do not cut through the mechanism question at all. Randomising or quasi-randomising the School Breakfast Program eliminates genetic confounding of treatment assignment while leaving the nutrition channel hopelessly entangled with income, attendance, take-up and stigma. Low genetic confounding and low mechanistic identification are different properties, and this literature routinely trades on the first to claim the second.

The pattern across the whole topic is the archive's standard signature: the more genetically informative the design, the smaller the effect — OR 2.08 observational, ~0.10 SD quasi-experimental, ~0.00 SD randomised.

Boundaries & what critics say

• This is not an argument against feeding children. Hunger is bad, food insecurity is bad, and Anderson's cost-effectiveness figure is genuinely excellent. The argument is against the nutrition mechanism claim, not against the program.
• "Deficiency" is a moving target and rich countries have it. Dunedin is not a poor country and those children were iodine deficient. The right posture is screening, not geographic assumption. Iron deficiency in menstruating adolescent girls is the obvious analogue.
• The strongest counter is that the null trials were underpowered or short. DOLAB II ran 16 weeks; Luan's trial was a secondary analysis of a study powered for obesity. Fair. But DOLAB II was larger than the trial it replicated and aimed at the responsive subgroup, and the burden of proof sits with the positive claim.
• Cochrane's LMIC math effect is real and GRADE-high. Someone can reasonably argue that 0.14 SD in genuinely food-insecure populations is the finding that matters most globally. It is — and it is also precisely the estimate that does not transport to a well-fed school.
• Frisvold and Imberman & Kugler are good papers. Nothing here says otherwise. The claim is that a quasi-experiment on program availability answers "should we run this program" and not "does nutrition raise achievement," and that where the identical contrast has been randomised the effect did not reproduce.
• Outcome-class discipline cuts against the deficiency findings too. Iron moves intelligence tests by 0.46–0.79 SD and moves school achievement by 0.06 SD (null). Whatever repletion is doing, it is not showing up where a school would measure it.

Practical guidance

• Screen for deficiency; do not supplement on suspicion. Iron (especially adolescent girls) and iodine are the two with real randomised support. A blood test costs less than a year of pills and tells you whether the pills can possibly work.
• Do not buy supplements, omega-3 capsules, or "brain food" for children who are already fed. Thirty-five RCTs and 19,343 children produced 0.09 SD on language. The DHA reading effect failed replication by its own authors.
• Do run a school meal program — and justify it correctly. At ≤$222 per 0.1 SD it competes with anything in this database. Sell it on food security, participation and cost-effectiveness. Staking it on "nutrition raises test scores" makes it vulnerable the moment someone reads the RCTs.
• If you must choose a margin, choose meal quality over meal timing. Anderson's healthy-vendor effect (0.036 SD, flat pre-trends, placebo-tested) is better identified as a food effect than any breakfast-delivery study, and moving breakfast into the classroom specifically has a null-to-negative randomised record.
• Ignore the breakfast-skipping literature when making decisions. It measures families, not meals.
• Stop managing "sugar highs." Blinded challenge experiments say the effect does not exist.

Open questions

• No high-income-country randomised evidence on school feeding and achievement essentially exists. Cochrane 2025 searched 17 databases and found one non-randomised breakfast-club study. That is the single largest gap in this topic.
• Nobody has run a screen-then-treat trial. Every micronutrient meta pools deficient and replete children and reports an average that describes neither. A trial that randomises supplementation within biochemically confirmed deficiency strata would settle the central question in one study.
• Whether iron/iodine repletion gains persist is unmeasured — follow-ups are 16–28 weeks. Fadeout has not been tested at all.
• Why iron moves intelligence tests but not school achievement is unexplained and important. It is the same dissociation the archive tracks elsewhere between test-score effects and taught-skill effects.
• No within-family or genetically informed study of childhood diet quality and achievement exists, which is why the entire observational half of this literature is uninterpretable rather than merely weak.