Does homeschooling raise achievement and attainment, or does it select the children who would have done well anyway?

Verdict

insufficient, and this is the topic where that verdict does the most work. Homeschooling has the widest gap in this archive between the confidence of the public claim and the quality of the evidence behind it. The claim — homeschooled children score at the 80th percentile, roughly 30 points above the national average — is real in the sense that those scores were really recorded. It is not evidence, because nothing was held constant.

Read the sentence the most-cited study actually contains. Rudner, whose 20,760-student dataset is the single largest ever assembled on homeschool achievement, wrote: "This study does not demonstrate that home schooling is superior to public or private schools. It should not be cited as evidence that our public schools are failing. It does not indicate that children will perform better academically if they are home schooled." The report that says this was funded by the Home School Legal Defense Association, whose president conceived it, and which distributed it to members of Congress. It has been cited as proof of exactly the three things its author disclaimed ever since.

The archive's premise cuts both ways here, and that is the point. If academic ability is substantially heritable, then a population defined by parental commitment, income, education and marital stability will produce high-scoring children under almost any delivery method — so the homeschool advantage is the expected observation under a null causal effect. But the same argument disqualifies the mirror claim. When a nationally representative survey finds homeschooled adolescents behind grade level, or a well-adjusted cohort finds them less likely to finish college, the leading explanation is reactive homeschooling: families withdraw children who are already struggling. Neither direction survives. Both are the same confound wearing different clothes.

What the evidence shows

The counterfactual test is the whole story, and nothing survives it. This was searched deliberately and exhaustively for this tranche, and the result is worth stating flatly: there is no RCT, no lottery, no regression discontinuity, no difference-in-differences on outcomes, no discordant-sibling or within-family design, no instrument, and no national register study of home education anywhere in the literature.

Each of the obvious openings was checked and each is empty:

• Regulatory variation. US states span mandatory annual testing plus portfolio review to no notification at all. Exactly two studies point at it, both authored by the movement's own research institute using its legal arm's regulation taxonomy: a state-aggregate SAT comparison with no covariates and no individual data, and an individual-score analysis of testing-service clients. Both report nulls, and Ray 2010's null is non-monotonic — low-regulation states scored lowest (z = 1.121) against medium 1.171 and high 1.131, with p-values of .045/.041/.084 declared non-significant against a pre-set α of .01. Neither design could detect a regulation effect if one existed. The wider literature (Hail, Levy, Stewart & Neeley, Bhatt) reaches the same weak-or-no correlation from equally weak designs.
• Bans as natural experiments. Germany, Sweden (2011), Spain, the Netherlands, Bulgaria. No outcome evaluation of any ban exists; recent German scholarship is, per Kunzman & Gaither, "entirely legal analysis." Sweden's home-schooling population was roughly a hundred families — a real policy shock with no cohort to measure.
• Sibling, register and lottery designs. Nothing at all. France's mandatory home-visit reports are being assembled into a population-level file that has not yet produced outcome estimates.
• The one causally identified paper in the entire literature has the wrong outcome. Bhatt's staggered-adoption analysis of home-schooling legalisation identifies an effect on take-up — a 1.4-percentage-point rise in the probability a young child is home-schooled — not on learning. The shock, the timing and the state administrative data all exist. Nobody has ever pointed that design at an achievement outcome. That is the single most actionable gap in this topic.

What is left is three designs that hold something fixed — Yu's matching on prior attainment (null), Chen's covariate-and-baseline-adjusted cohort (negative on attainment), and the cyber-charter switcher designs (large negative, but a different treatment) — and, beyond them, self-selected families compared to a national norming table.

The advocacy side concedes the point in its own words. Ray's 2026 systematic review, with the whole literature in front of him, recommends the field advance through "well-controlled non-experimental design research." He does not claim a causal design exists, because there isn't one.

Two internal drifts inside the advocacy corpus are worth more than its headline. First, the same author, using the same testing-service design, reported the mean at the 84th-89th percentile in 2010 and at the 75th-79th percentile in 2026 — a 5-10 percentile slide with no change in method, which is what one expects as a movement grows beyond its founding demographic. Second, the 2026 vote count is 62% of achievement studies positive: even scored by its most sympathetic reader, using a method with no design screen, close to four studies in ten fail to find the advantage.

The most useful single number in the literature is Yu's r = .21 versus .51. High-school GPA predicts first-year college GPA less than half as well for homeschooled students as for matched traditional students, while the SAT predicts about equally well for both. That is a direct measurement of the archive's measure-type rule inside this topic: the parent-assigned grade carries much less information than the externally scored test. Any homeschool achievement claim resting on parent-administered or parent-scored assessment should be discounted on that basis alone — and Rudner's tests were administered largely by parents in their own homes.

Where a signal does look real, it is about content, not about home delivery. Gaither documents the same asymmetry across at least six independent samples spanning three decades — Frost & Morris, Wartes, Ray's own 1997 data, Rudner, Belfield's SAT analysis, Qaqish's ACT analysis, Coleman's state-data review: homeschooled students look relatively strong on verbal measures and relatively weak on mathematics. Guterman & Neuman's matched Israeli sample adds the mechanism-shaped version of the same thing: broader general knowledge, equal listening comprehension, but lower phonological awareness and lower reading comprehension. That is the profile of a home rich in language and content that does not systematically teach decoding or drill arithmetic — which is a statement about what gets taught, and it is actionable in a way the aggregate percentile is not.

Belfield's SAT decomposition is the cleanest version of that finding and it also disposes of the headline. Across every self-identified home-schooled 2001 SAT taker: raw total 1093.1 against 1012.6 for public school — the ~0.4 SD figure the movement quotes — but 0.15 SD below private-independent schools at 1123.8, which nobody quotes. The premium is almost entirely Verbal (566.6 against 502.6) and almost absent in Math (526.5 against 510.1); compared to their own predicted scores, home-schoolers came in fractionally below prediction on Math. And once family background is adjusted, the four sectors converge: home 1054.5, public 1021.1, private-religious 1050.5, private-independent 1064.4. The home-school advantage over religious schooling falls to roughly zero. Belfield also reports the finding that matters most for a family deciding: the SES gradient is steeper under home education than in any other sector. Family background does more of the work, not less.

On attainment the correlational record leans negative, and it does so from probability samples rather than recruited ones. The Cardus surveys are random panels, not alumni networks recruited through movement newsletters, and they find home-school graduates over-represented at open-admission universities, essentially absent from Carnegie R1/R2 institutions, with no years-of-education advantage, about one year less postsecondary education net of a substantial control set, lower odds of an advanced degree, and lower income. Chen's nurses'-children cohort reaches the same place (RR 0.77 for a college degree) from an entirely different dataset with an entirely different design. That convergence is the strongest signal in the topic pointing against the popular claim — and it is still observational, still compatible with reactive withdrawal, and still not a counterfactual.

Three internal analyses inside Rudner's own confounded data are worth more than his headline. A parent holding a state teaching certificate made no difference (F=2.9, n.s., controlling for grade and parent education). Enrolment in a packaged full-service curriculum made no difference (F=0.24, n.s.). Parental education made an enormous difference (F=566.4, p<.01) — 83rd-98th percentile for children of two graduates, mostly 66th-69th for children of non-graduates. Selection cannot explain within-sample contrasts nearly as well as it explains the headline, so these are the more informative rows in the paper. They are developed further in parent-delivered instruction.

Hereditarian-lens assessment

Risk: high, and unusually, the selection is measurable rather than hypothesised.

Rudner's own text describes it: family income and education "well above national averages," married parents, a stay-at-home mother, and 23.6% of students with at least one parent who has held a state teaching certificate. Ray's 2010 sample was 95% Christian, 91.7% white, 97.7% married, 80% stay-at-home mothers, 45.9% earning over $80,000. These are not populations that would produce average test scores in any setting. The archive's adoption evidence puts a ceiling on how much of the parent-child association is rearing at all: a college-educated mother raises an adoptee's college graduation by 7 points versus 26 for her biological child.

Three things follow, and the third is the one usually missed:

• The raw gap is uninformative. Yu's design demonstrates this rather than assuming it: a 0.52 SD raw SAT advantage becomes an exact null once you match on prior attainment and demographics.
• The anti-homeschooling findings are equally confounded. Chen's 23% attainment penalty and Green-Hennessy's threefold grade-retardation odds are observational. Reactive withdrawal of a struggling child would produce both. Chen's E-values (1.54-2.04) say the associations are not trivially explained away — but an unmeasured confounder of that strength is entirely plausible when the exposure is why a family left school.
• The within-homeschool gradients are the least confounded thing here, and they point at the parent. Belfield found greater variance in SAT scores by family background among homeschoolers than among schooled students; Boulter's longitudinal sample, where parents averaged only 13 years of education, showed achievement declining the longer children were homeschooled. School compresses the effect of home; removing school widens it. That is the single most decision-relevant implication of the hereditarian frame for this topic, and it is not a claim about the average homeschooler.

Boundaries & what critics say

• There is no natural experiment, and it is not for want of looking. Homeschooling is chosen, never assigned; the countries with population registers good enough for a clean design (Sweden, Germany) are the ones where the practice is effectively banned, so the shock exists and the cohort does not; and the one US legalisation wave that has been exploited for identification measured take-up rather than learning. Carlson documents that state regimes vary from standardized testing plus portfolio review to no notification at all — a natural-experiment opportunity that nobody independent has taken. This is the largest single gap in the topic.
• The regulation nulls are quoted far beyond what they can carry. "Regulation doesn't improve home-school outcomes" rests on two studies by the movement's own research institute, using the regulation taxonomy of the movement's own legal arm, on samples of families who volunteered to be tested. One has no individual-level data at all. Both are underpowered for the question, and Ray 2010's own pattern is non-monotonic in the wrong direction for the argument it is used to support. This is a null from designs that could not have found the effect, which is not the same as evidence of no effect.
• The virtual-schooling evidence is the closest thing to a causal estimate of moving school home, and it is not homeschooling. Three independent datasets and three methodologies agree on sign and magnitude class: CREDO's national matched-twin study gives −0.25 SD math and −0.10 SD reading per year, negative in every racial-ethnic subgroup, negative in essentially every state, and negative against brick-and-mortar charter comparisons as well as district ones — which rules out the charter-versus-online confound. Its within-chooser analysis, comparing eventual switchers to themselves in their prior traditional-school year, is larger still at −0.17 SD reading and −0.34 SD math across roughly 120,000 students. Cordes's Pennsylvania estimates (−9.5 pp graduation, −16.8 pp college enrolment) and Fitzpatrick's within-student Indiana switchers replicate the sign. Tellingly, CREDO's leavers who returned to traditional schools posted +0.55 SD in math in their second year back. But in a cyber charter the curriculum is a vendor's, the teacher is remote, and the parent is a "learning coach" rather than an instructor — and enrolment is often reactive, which biases against virtual schooling rather than for it. Read this as a boundary marker on unsupervised home delivery of somebody else's programme, not as a verdict on parent-led education.
• Engzell's COVID estimate is a genuine natural experiment on a caricature. A national closure assigns home learning to everybody at once, which strips out family selection completely — and produces −0.08 SD in eight weeks, meaning children made essentially no progress at home, with losses 60% larger in less-educated households. That isolates the delivery context from the selection. But the exposure was involuntary, unplanned, uncurriculumed, and delivered by parents who were simultaneously trying to work. Under this archive's own rule that a null on a caricature counts for nothing, it cannot be used to overturn chosen, prepared, structured home education. What it does establish is that home delivery is not automatic: the parent-supplied structure is load-bearing.
• "Homeschooling" is not a single exposure and the pooled estimate may be meaningless. Cheng & Hamlin show that most contemporary US homeschool families buy co-op instruction, private tutoring, online provision or part-time brick-and-mortar schooling; the fully self-contained arrangement is a shrinking minority concentrated among White, less-educated, Southern families. Meanwhile Martin-Chang finds structured and unstructured homeschoolers separated by 1.3 to 4.2 grade levels — a spread far larger than any homeschool-versus-school difference anyone has reported. A study that pools these is averaging across a wider range than it is comparing.
• The advocacy literature is not merely biased, it is structurally so. Ray is president of NHERI, which produces most of the primary studies his reviews review; the samples are recruited by HSLDA with, in Gaither's account, "promises made that the research will be used for homeschooling advocacy"; the tests are proctored by parents; the scores are self-reported. Ray's reviews then vote-count those studies by the direction of their result. Recorded as developer-led. The value of keeping it in the archive is that this is the evidence base most readers will encounter.
• Kunzman & Gaither's 2020 comprehensive survey could not be obtained — the journal host does not resolve, there is no DOI, and no open-access copy exists in ERIC, Unpaywall, Semantic Scholar or IA Scholar. Recorded as unresolved/inaccessible. Gaither's 2017 single-author review, drawing on the same underlying survey, was obtained in full and carries the load.

Practical guidance

• Ignore the percentile claims in both directions. "Homeschoolers score in the 80th percentile" and "homeschoolers are less likely to finish college" are the same kind of statement about the same kind of data. Neither is a reason to homeschool or not to.
• If you homeschool, structure is the decision that has evidence behind it. The largest measured contrast inside this literature is not homeschool versus school; it is structured versus unstructured homeschooling, at 1.3-4.2 grade levels. Unschooling is the one arrangement with a recorded null-to-negative result against an ordinary public-school comparison.
• Teach decoding explicitly and protect mathematics time. The verbal-strong/maths-weak profile replicates across six independent samples and three decades, and the one study that decomposed reading found the gap sitting in phonological awareness. These are the two most predictable failure modes of home delivery and both are fixable with the archive's existing phonics and math fact fluency recommendations.
• Do not use parent-assigned grades to know how it is going. They predict later performance at less than half the validity of an external test. Buy an independent standardized assessment, and do not proctor it yourself.
• Budget the real cost. It is one adult's full-time labour, not curriculum materials — Rudner's families spent little on materials and 62% of the mothers in Martin-Chang's homeschool group had left paid work. Any comparison against tutoring or private school that ignores this is not a cost comparison.
• The socialisation question is a separate decision with a separate answer. See homeschool socialisation: total social contact is not the problem, peer contact and friendship closeness are, and no amount of achievement evidence speaks to it.
• Treat outside activity as a requirement, not a bonus. The one subgroup that nationally representative data flag as at risk is homeschoolers without religious or community ties, who were three times more likely to be behind grade level and 2.5 times more likely to report no extracurricular activity at all.

Open questions

• The obvious study has never been done, and the instrument is already sitting there. State home-schooling legalisation was staggered across decades, the shock is exogenous to any individual family, and state administrative achievement data exist for the same period. Exactly one paper has used that design and its outcome variable was take-up (+1.4 pp), not learning. Pointing the same identification at an achievement outcome would answer more than the entire existing corpus.
• A sibling design would be feasible and nobody has run one. Many families home-school some children and not others, or start mid-childhood. Within-family comparison removes the confounding that makes this literature useless, and no study in the corpus attempts it.
• Nothing usable exists on adult outcomes beyond a single cohort and two Cardus waves. Chen's GUTS analysis is one cohort of one occupational group's children with n=187 home-schooled; Cardus publishes no coefficient tables, standard errors or confidence intervals, so its "about one year less" cannot be interrogated. Earnings are essentially unmeasured.
• The structured/unstructured contrast is the field's biggest measured effect and rests on twelve children in one exploratory subgroup. It deserves a properly powered, prospectively defined replication more than anything else in this topic.
• Why the mathematics deficit? The pattern is well replicated and the explanations are all speculative — less priority, less parental confidence, less drill, more conversational instruction. No study has measured instructional time by subject in homeschool households against outcomes.
• What are the outcomes of the modal contemporary arrangement? Most homeschoolers now use co-ops, tutors, online providers and part-time enrolment. Essentially all of the outcome evidence predates that shift or ignores it.