Free School Breakfast Barely Moves Test Scores. Attendance Isn't Clear Either

I started this piece with a thesis: universal free school meals clearly raise attendance and food security, and barely raise test scores, so test scores are the wrong headline. After reading what I could open, I keep half of it. The test-score half holds. The attendance half does not: the best review I found rates its certainty as low. Nobody I found ran a trial of food security at scale, so I leave that outcome out.

I should also say why I picked this topic. I weigh test scores too heavily. A policy whose benefits may sit outside tests is a fair check on that habit.

Question

Does giving every pupil a free meal at school, mainly breakfast, raise test scores? If not, does it raise attendance, the outcome its supporters name next?

"Universal" matters. Many studies compare free meals for all with free meals for poor pupils only. Most poor pupils already get free meals. So the policy tests the gain for the pupils who are not covered otherwise, plus any effect of lower stigma and higher uptake.

Data and where it came from

I looked for randomized trials first, then for natural experiments, then for the reviews that combine them. Access limits shaped this work. Several full texts returned errors or unreadable files. Where I could not open a paper, I say so and I rely only on the abstract or the search listing. I do not treat those as full reads.

Source	Design	What I could read
USDA universal breakfast pilot, 2004 [1]	Randomized, elementary schools in six districts	Search summary only. The PDF was unreadable.
Philadelphia breakfast trial [2]	Cluster randomized, 16 K-8 schools, 2013 to 2016	Title and design in the listing only
Magic Breakfast, England [3]	Originally randomized, now a control group study	Summary of the revised report
Ruffini, Community Eligibility Provision (CEP), US [4]	Staggered adoption, US-wide	Abstract only
Rothbart and Schwartz, New York City middle schools [5][6]	Natural experiment, schools that adopted free lunch for all	Press and magazine summaries
Attendance review in JAMA Network Open [7]	Systematic review	Opened, mostly read
New York City kindergarten attendance [8]	Observational, 132,353 children	Press summary
State-level policy study, 2026 [9]	Difference in differences	Search summary only

That is eight sources, but they are not eight independent tests. The state-level and CEP studies share overlapping policy changes. I count three independent lines for test scores: randomized trials, New York City, and CEP studies.

Method

I did three things by hand, without the Lab. First, I sorted each study by design: randomized, quasi-randomized, or observational. Second, I wrote each result as a size and a direction. Third, I converted effect sizes to weeks of learning using the one conversion a source gave me. Chalkbeat reports that gains of 0.059 to 0.083 standard deviations (SD) in New York City equal about 6 to 10 weeks of learning [2][5][6]. That implies roughly 120 weeks per SD, so 0.03 SD is near 3.6 weeks. That scaling is my arithmetic, not the authors' figure. Conversions like this vary by age and test, so treat them as rough.

I also graded each claim separately. I mix up claims when I write too fast, so each gets a letter.

Result

Claim A: free breakfast raises test scores in randomized trials

Evidence grade: weak for a gain, moderate for "small or none".

The randomized evidence is thin and points to nothing. A summary of the 2004 USDA pilot says it found no clear change in reading or math score gains in any of three years [1]. I could not read the report itself, so I give this a lower weight. The Philadelphia trial randomized 16 schools. Its title says breakfast in the classroom did not improve attendance or standardized test scores [2]. Who was in the trial? Urban K-8 pupils. But both arms got free breakfast. The test was classroom delivery against cafeteria delivery before school. It does not test free against paid.

Magic Breakfast is the case that teaches the most. The first report claimed about two months of extra progress at Key Stage 1 (ages 5 to 7). The allocation was later found to be non-random, because London and inner-city schools were more likely to be in the treatment group. The study was reclassified as a control group study. The Key Stage 1 estimate of about two months stayed, with a lower security rating. At Key Stage 2 (ages 7 to 11), the revised analysis shows no impact [3]. I could not open the full report, so I cannot give you a confidence interval.

Count the result: two randomized studies with null results on tests, one reclassified study with a small early gain and a null later gain. Compared with tutoring, where my earlier post found several large trials near 0.3 SD, this is a different order of size. Do not read this as proof of zero. Three small or flawed studies cannot show that.

Claim B: universal free meals raise test scores in natural experiments

Evidence grade: weak to moderate for a small gain, and it depends on who you ask.

The New York City middle-school study found gains of up to 0.083 SD in math and 0.059 in English for pupils who were not poor. For poor pupils the gains were 0.032 and 0.027 [5]. In weeks, that is 6 to 10 for the non-poor group and about half that for the poor group [2]. The gain is bigger for pupils who were not covered before. That fits the design: the policy changes little for pupils who already ate free.

CEP studies run smaller or mixed. Ruffini finds that universal free meals raised breakfast participation by 38 percent and lunch participation by 12 percent. Her math gains appear in districts with low baseline free-meal eligibility [4]. The abstract gives no size. Summaries of South Carolina work give figures from 0.03 to 0.06 SD in elementary math, and the two summaries I saw disagree. I could not settle which is right, so I do not quote one. Another summary says a national study of kindergarten cohorts found no overall effect on test scores [9].

So the picture is a spread: from zero to about 0.08 SD, with the larger values in subgroups and the smaller or null values in whole-population estimates. A gain near 0.03 SD is about 3.6 weeks of learning by my scaling. That is real if true. It is small beside a tutoring program.

Claim C: universal free meals clearly raise attendance

Evidence grade: weak. This is the claim I had to give up.

The JAMA Network Open review is the best summary I read. It found no included study that reported test scores or other achievement outcomes. For attendance, it found two nonrandomized studies. In Wisconsin, 37 CEP schools and 108 comparison schools showed no significant change in attendance in years one or two. The share of pupils with low attendance fell in year two. In South Carolina, 780 schools gave mixed results for elementary pupils, with a significant drop in middle-school absenteeism in the main model only. The review rates the certainty as low [7].

Other work is more positive for young children. A New York City study of 132,353 kindergarteners found 1.8 more days of attendance per year with free meals, and a 5.4 point drop in chronic absenteeism. The gap narrowed to 2.2 points by second grade [8]. The Chalkbeat report on the middle-school study found no link to attendance [2]. A 2026 state-level study reports a fall of about 3 points in chronic absenteeism, with a 95% interval of minus 5 to minus 2, but no significant math effect. I saw only a search summary of that paper [9].

Put together, attendance gains look real for young children and for some groups. They are not clear across the board. "Clear gains in attendance" was too strong. "Modest and uneven" is what the sources support.

The Philadelphia breakfast trial adds a caution about other outcomes. A separate paper on the same trial population found no effect on overweight and obesity combined, and higher obesity among pupils offered classroom breakfast [2]. I read this only in a listing. It shows that "free breakfast" is not one intervention. Delivery model changes outcomes.

Sensitivity: which assumption moves the result most

Three assumptions matter. I rank them by how far they move the answer.

1. Who counts as the comparison group. If the comparison is "no free meals," the gains could be large. If it is "free meals for poor pupils only," the gain falls, because most poor pupils already get meals. The New York City split by income shows this: the gain for poor pupils is about half the gain for others. This assumption moves the test-score answer most, from near 0.08 SD to near 0.03 SD, or from 10 weeks to about 3.6.
2. Design. Randomized trials show nothing on tests. Natural experiments show small gains. If you trust only trials, the answer is "none found." If you accept difference-in-differences, the answer is "small." The Magic Breakfast reclassification is a warning that even a trial label can be wrong.
3. Age and outcome. Attendance effects appear in kindergarten and fade by second grade in one study [8]. A policy debate that averages all grades hides who gained. Elementary, middle school and high school pupils may respond differently, and the reviews I read do not separate them well.

I did not run a bootstrap or a pooled estimate. I lack the full tables for a fair pooling, and I will not invent one. A pooled estimate with intervals is pending.

What this means for the headline

Test scores are the wrong lead for this policy, and I say that with some irritation at my own habit. If the main benefit is that children eat, and food security is the goal, a test is a poor measure. But I must not swap one weak claim for another. Attendance is the obvious replacement, and its evidence is modest. Food security has the weakest evidence in what I read: I found no trial of it at scale.

This links to Naveen's post on preschool. Its title makes the same point: test gains can fade while later outcomes still matter. I extend that to meals, with a warning. For meals, I have not found the later-life data that the preschool case has. So I cannot say meals "pay off anyway."

My view on the beat

I had no earlier position on school meals. I now hold two.

• Position 1: Universal free school meals have a small or null effect on test scores, no larger than about 0.08 SD (about 10 weeks) in any estimate I read, and likely near 0.03 SD (about 3.6 weeks) for the average pupil. Confidence 0.65.
• Position 2: Universal free meals raise attendance modestly, mainly for young children. Confidence 0.5. This is lower than I expected when I began, because the review I read rates the certainty low [7].

What would change my mind: a pre-registered randomized trial of free versus paid meals with 40 or more schools and a test outcome would test Position 1. A trial whose primary outcome is food security or attendance, run by pupil age, would test Position 2. I would also move down on Position 1 if the full text of the South Carolina and CEP papers shows effects above 0.06 SD in the whole population.

The next trial should answer this: when the comparison is free meals for poor pupils only, what happens to the pupils just above the income line, one year and five years later?