The parent as instructor — what survives when the person teaching is the person who raised the child

Verdict

mixed, and the mixture is not noise — it is a boundary. The same parent, with the same child, in the same half hour, produces d = 1.15, d = 0.51, or nothing, depending only on how specifically the task is defined. Sénéchal & Young's meta-analysis of intervention studies separates three levels of instructional specificity and finds a monotonic gradient: parents tutoring a named literacy skill 1.15; parents listening to a child read 0.51; parents reading to a child 0.18 and not significant. Nothing about the parent changed between those three cells.

That gradient is the answer to "how do methods degrade when a parent delivers them". They do not degrade because the deliverer is a parent. They degrade because home delivery is where instructional specification is most likely to be missing — and when it is supplied, a parent gets a real effect.

But do not take d = 1.15 home with you. The 2025 meta-analysis that separates group designs from single-case designs for exactly this intervention class — family members teaching literacy to school-aged children — gets g = 0.36 from the group designs and g = 1.50 from the single-case designs. Four times apart, same activity. The archive's benchmark says large effects from small studies are a red flag rather than a triumph, and here that rule is vindicated inside the literature itself. The honest end-of-programme number for a parent teaching a school-aged child to read is about a third of a standard deviation — code-focused skills 0.28, meaning-focused 0.41.

And then it fades, faster than almost anything else in this archive. Two independent meta-analyses of family literacy programmes measure a follow-up: van Steensel's 30 evaluations give d = 0.18 overall, 0.11 in the randomized subset, and 0.04 (not significant) at follow-up; Fikrat-Wevers's 48 studies of low-SES families give 0.50 at posttest and 0.16 at follow-up, with comprehension-related skills collapsing almost entirely (0.51 → 0.09) and only code-related skills holding any of it (0.48 → 0.22). Half of even the posttest number is measurement: researcher-made instruments returned 0.84 against 0.37 for existing standardized tests. Stack the discounts — randomize, measure with an instrument you did not write, and come back three months later — and the effect of a parent-run literacy programme is somewhere between 0.05 and 0.15 SD. That is not zero. It is also not a school year.

And the ceiling is low against the alternatives. In the only meta-analysis that puts tutor types on a common scale, parents are at the bottom. Nickow, Oreopoulos and Quan's tutor-type panel — the working-paper version, which is the only one that reports a parent cell — gives teachers 0.50, paraprofessionals 0.40, non-professional volunteers 0.21 — and parents 0.23 with a standard error of 0.114 across 11 studies, significant only at the 10% level, drawn from a literature with no large-scale evaluation of parent tutoring at all and in which, in the authors' own accounting, no parent-tutoring study met their low-bias quality criteria. Their own summary is blunt: "The experimental parent tutoring research is still too thin and fragmented for consistent lessons." Fifteen years of additional studies have not fixed it — a 2023 audit of 30 parent-tutoring studies against the Council for Exceptional Children standards rates the practice "potentially evidence-based" for kindergarten through grade 3 and "insufficient evidence" for grades 4 through 12. Parents are the cheapest tutor and roughly the least-well-evidenced one.

What the evidence shows

Does parental subject knowledge or training move outcomes? The three answers do not agree, and the disagreement is informative.

• Credentials: no. Inside Rudner's 20,760-student homeschool sample, controlling for grade and parent education, children whose parent held a state teaching certificate scored no differently (F=2.9, n.s.) — percentile differences of −2 to +5, in both directions. This is the homeschool replication of the archive's teacher-quality finding that credentials do not identify effective instructors.
• Education: yes, hugely — in observational data. In the same sample, children of two college graduates sat at the 83rd-98th percentile and children of non-graduates mostly at the 66th-69th (F=566.4, p<.01). Medlin found mother's education predicting achievement among homeschoolers; Boulter's longitudinal sample, where parents averaged 13 years of schooling, showed scores declining with more years of home education.
• Education, causally: close to nothing, and this is now measured directly rather than inferred. Three grade-B designs attack the parental-schooling → child-outcome link and all three collapse it. Black, Devereux and Salvanes instrument Norwegian parents' schooling with a compulsory-schooling reform across 286,137 mother-child pairs: OLS says 0.15 years per parental year, every full-sample 2SLS estimate is insignificant, and where the instrument bites, fathers are ~0 and the only surviving effect is mothers on sons, 0.105 (SE 0.039). Holmlund, Lindahl and Plug run twins, adoptees and IV on the same Swedish data and every design lands below an OLS of 0.28/0.23: twin-parent fixed effects give mothers −0.001 and fathers 0.120; adoptees 0.01-0.11; IV mothers ~0.06. Behrman and Rosenzweig, differencing within pairs of identical-twin mothers, get a negative coefficient. The observational F=566 inside Rudner's homeschool sample is heredity and assortative mating, not a teaching effect.
• The adoption evidence says the same thing from the other direction, and it is the same argument this archive makes at length in the early home environment, where the correlational parenting literature is shown to be measuring heredity. A college-educated mother raises an adoptee's college graduation by 7 points against 26 for her biological child — roughly three-quarters of ordinary parent-child transmission is not causal rearing. But keep one caveat that matters here and nowhere else in the archive: a homeschooling parent who cannot do the mathematics cannot teach the mathematics. The causal near-null on parental years of schooling is not the same claim as a null on parental command of the content being taught, and no study separates them.
• Training: the training demonstrably changes the parent, and the effect on the child is much smaller. In Heidlage's 25 RCTs the largest single effect in the whole meta-analysis is on the ADULT — parent use of language-facilitating behaviours, g = 1.20 — while the child effects are expressive vocabulary 0.42, expressive language 0.27, receptive vocabulary 0.18 (n.s.) and receptive language 0.07 (n.s.). A trained parent reliably does the thing; the thing moves the trained target and not the untrained capacity. Roberts & Kaiser's range (−0.15 to 0.82) swings on the comparison group and on whether the parent or an independent observer scored the child.
• How much training and supervision is required is genuinely unresolved, and the intuitive answer is not supported. Sénéchal & Young found that providing supportive feedback to parents during the intervention did not produce better child outcomes (8 studies with, 6 without), and that less training was associated with larger effects — a comparison they decline to interpret because it is confounded with intervention type. Kupzyk's audit names the reason nobody can settle this: across 30 parent-tutoring studies, training and fidelity are described inadequately.

Who trains the parent is a measured null, twice. This is the most practically useful finding in the topic and it is easy to miss because it is a negative. van Steensel's meta found no significant moderators of family-literacy effects at all — not programme duration, not home visits versus group meetings, not book provision, not at-risk status, and not trainer type (professionals d = 0.21, semiprofessionals 0.18, both 0.12, Q = 0.59 n.s.; and identical at 0.16 within at-risk families, Q = 0.00). Fikrat-Wevers, on a different corpus a decade later, got the same answer (professionals 0.65, paraprofessionals 0.40, Q(3) = 5.39, n.s.). Two independent tests, same null. The expensive part of these programmes is not the part that works. What the parent is asked to do carries the effect; who briefs them does not.

Parent-delivered mathematics is not established, and its flagship trial is null. The Bedtime Math study is cited as evidence that a parent-delivered maths app buys almost three months of extra achievement. Read as a trial, it does not show that: there is no overall intent-to-treat main effect reported anywhere in the paper, the headline lives in a median-split subgroup of high-maths-anxious parents (b = 5.25, P = .048) licensed by an interaction that was P = 0.06 — a fact disclosed in the authors' 2016 Response rather than in the original Report — and the rest rests on a non-randomized dose-response. Frank's reanalysis of the authors' own posted data found no significant intervention effect and no condition-by-time interaction. This is the archive's standard pattern arriving on schedule.

What does survive in parent-delivered maths is a readiness constraint rather than an effect size. Gibson, Gunderson and Levine randomized which number book parents read: children given small-number (1-3) books gained 0.97 knower-levels against 0.23 for controls, while the same parents reading large-number (4-6) books produced nothing — and among children who had already mastered small numbers, the pattern reversed. Same parents, same activity, same amount of time. The instruction has to be pitched at where the child actually is, which is the one thing a parent teaching one child is structurally better placed to do than a teacher with thirty.

More parental help is not better, and the one form of involvement that looks harmful is the one most parents actually do. Barger's meta of 448 studies covering 480,830 families finds involvement broadly and weakly positive (r = .13 to .23 for academic adjustment) with a single exception: homework assistance, r = −.15. Fernández-Alonso's pooling of 180 effects across PISA 2009, 2012, 2015 and 2018 finds d = 0.23 favouring students who get less family help, invariant across subject and stable across a decade, though not across culture (Europe 0.30 against Southeast Asia 0.09). Maloney adds the mechanism-shaped version: maths-anxious parents who frequently helped with homework had children who learned less maths across the year and ended it more maths-anxious (interaction F(1,430) = 4.59, p = .033) — and the effect was maths-specific (reading F = 0.114, p = .736) and survived controls for parental maths knowledge, so it is anxiety transmission rather than incompetence. All three are observational and reverse causation is live — struggling children get more help — so none of this is evidence that helping causes harm. What it does establish is that undirected "help with the homework" has no positive evidence behind it whatsoever, in the largest samples anybody has assembled.

Undirected involvement is where the field collapses. See & Gorard screened 4,898 reports, found 127 attempted evaluations, and concluded that none was large and robust; the 121 weak ones split almost evenly between claimed success and ineffective-or-harmful. Nye's Campbell review of 19 RCTs reports d=0.45, but from trials with a median length of 11 weeks and mixed measure types — the efficacy-phase profile this archive repeatedly watches deflate. Avvisati's cluster RCT is the cleanest test available and splits the difference in the most instructive way: the programme genuinely raised parental involvement, truancy and disciplinary sanctions improved by about 15% of an SD, and test scores did not move at all.

The one thing parents do better than tutors is not teaching. Bergman's biweekly missed-assignment alerts produced +0.19-0.20 SD on GPA and +0.21 SD on maths for a few text messages, and York, Loeb & Doss's parent text nudges produced +0.11 SD against an active control. Both work through monitoring and accountability rather than instruction. Per hour of parental time, this is the highest-return use of a parent found anywhere in this archive — and it requires no subject knowledge whatsoever.

Hereditarian-lens assessment

Risk: medium, and the topic splits cleanly into a low-risk half and a high-risk half.

Low risk: everything with random assignment. Sénéchal & Young's gradient, Nye's RCTs, Heidlage's 25 trials, Avvisati, Bergman, York/Loeb/Doss, and the tutor-type panel all compare families that were equally selected and differ only in what they were asked to do. The parent's genotype is constant across arms by construction. When this archive says "instruction works", this is the kind of evidence it means.

High risk: everything about which parents do better. The parental-education gradient inside homeschooling, the correlation between home literacy environment and reading, the finding that homeschooler achievement varies more by family background than schooled achievement does — all of it is the classic passive gene-environment correlation. Plomin's replicated finding is the summary statistic: about two-thirds of parenting-to-child correlations are genetically mediated, and measures of "the environment" are themselves heritable at h²≈0.27.

The interaction is the interesting part, and it points the opposite way from the usual reassurance. School is a floor. Remove it and the variance attributable to the family goes up — which is exactly what Belfield observed (more score variance by family background among homeschoolers) and what Boulter's low-education sample shows (decline with more years at home). Under the archive's own framing, good universal instruction increases the heritability of a skill by removing environmental bottlenecks; home delivery does the reverse, reintroducing the bottleneck. This does not say homeschooling is worse on average. It says the spread is wider, and that the parent's own capability is doing more of the work than it would in a school.

Boundaries & what critics say

• The parent-tutor estimate rests on 11 small studies and could easily be wrong. 0.23 with a standard error of 0.114 is not a precise number, and Nickow et al. explicitly note there were no large-scale evaluations of parent tutoring in their sample and that the largest one they had (a Hong Kong preschool paired-reading programme, just under 200 children) came in near their overall pooled effect of about a third of an SD. They also note the low parent and volunteer estimates are driven mainly by reading programmes, since nearly all maths tutoring used paraprofessionals. It is entirely consistent with the data that parent tutoring is roughly as good as paraprofessional tutoring and has simply never been evaluated properly — and Dahl-Leonard's independent g = 0.36 for family-implemented literacy instruction sits comfortably above the 0.23. What is not consistent with the data is treating parent delivery as the premium option.
• The parent arm is confounded with the worst delivery conditions. 10 of the 11 parent-tutoring studies ran after school, and essentially all ran 1:1 at home. The tutoring literature's two most robust structural findings are that in-school beats after-school by roughly 2x and that scheduled daily dosage beats opt-in. Parents are being scored under precisely the conditions that depress every tutor type. Nobody has tested a parent tutoring in a protected, scheduled, in-school-equivalent slot — which is exactly the condition a homeschooling family is uniquely able to create.
• d = 1.15 is a red flag by this archive's own rules, not a triumph. It rests on seven small studies; Sénéchal & Young's own effects were heterogeneous (0.07 to 2.02) and their own text warns that "caution should be used in interpreting the results"; and within their data standardized tests produced smaller effects than researcher-designed ones — the archive's 2x measure-type inflation showing up inside the meta-analysis. Dahl-Leonard's design-separated g = 0.36 is the number to plan with. The ranking across the three activity types is the durable finding; the top cell's magnitude is not.
• Nye and See & Gorard cannot both be right about the state of the field. Nye pooled 19 short RCTs and got 0.45; See & Gorard graded the whole literature including those trials and found nothing robust. This file sides with neither and reports both, because the disagreement is about how much weight small short trials deserve — and that is a weighting question the archive deliberately leaves tunable.
• The parent-implemented language literature sits below the archive's age floor (roughly 18-60 months) and concerns children with language impairment. It is recorded for the delivery-agent question — can a trained parent execute a protocol, and what moves when they do — not as an age-relevant effect size.
• Shared reading is the most over-recommended parent activity in existence — unless you change what happens during it. Plain shared reading against a passive control looks fine (g=0.23); against any other adult-attention activity it is g=0.03, and Sénéchal & Young's three read-to-child studies gave d=0.18, not significant. But dialogic reading — the parent prompting and expanding rather than the child listening — buys d=0.59 on expressive vocabulary against ordinary shared reading as the active control (Mol, k=9, n=322). That is the same specificity gradient again, inside the one activity everybody already does. The catch is in Mol's moderators and it is the archive's usual disappointment: the effect shrinks for 4-5-year-olds and for children already at risk of language and literacy difficulty. The authors' own conclusion is that dialogic reading changes home literacy for families of 2-3-year-olds "but not those of families with children at greatest risk for school failure" — it works where it is least needed.
• The high outlier of this literature is Erion's +0.55, and it should make you more suspicious, not less. A 2006 meta of 37 parent-tutoring studies puts parent delivery above the pooled tutoring effect of 0.288 and far above the nonprofessional cell. Given that the same corpus fails a CEC quality audit for grades 4-12, that the design-separated estimate is 0.36, and that every better-identified subset in this file comes in lower, the 0.55 is best read as what a small-n, short-horizon, researcher-measured literature produces before anyone controls for those things.
• This topic does not settle whether a parent can deliver a full curriculum. Every clean estimate here is of a parent delivering one specified skill for a bounded period. Extrapolating from "a trained parent can raise decoding over ten weeks" to "a parent can run all subjects for twelve years" is exactly the over-application the dose field exists to prevent. See homeschooling outcomes for what happens when people try.

Practical guidance

• Buy the script, not the sentiment. Every positive result in this file involves a parent being told precisely what to do and given the materials and the method. Every null involves a parent being encouraged to be involved.
• Plan on g ≈ 0.3 at the end and g ≈ 0.1 later, not d ≈ 1. The design-separated estimate for a family member teaching literacy to a school-aged child is 0.36 (0.28 code-focused, 0.41 meaning-focused); randomize it and it drops toward 0.11; come back at follow-up and it is 0.04 to 0.16. Worthwhile for a free intervention. Not a substitute for a school year, and not durable without re-dosing.
• Do not pay for the expensive trainer. Who trains the parent is a null in two independent meta-analyses, including within at-risk families where it was identical to two decimal places. Spend the money on the materials and the specification instead.
• Do not treat "help with the homework" as instruction. It is the single involvement type negatively associated with achievement in the two largest datasets ever assembled on the question, and for a maths-anxious parent the association with maths learning is negative and specific. If the parent's own relationship with the subject is bad, outsource that subject.
• If you are going to read to a young child anyway, make it dialogic. Prompting and expanding rather than reading at them is worth d = 0.59 on expressive vocabulary against ordinary shared reading — one of the few free upgrades in this archive. Expect less of it for 4-5-year-olds and for children already behind.
• Pitch the content at where the child is, not where the curriculum says. The same parents reading the same kind of book produced +0.97 knower-levels with content matched to the child's level and nothing with content one step too far ahead. Fine-grained matching is the one structural advantage parent delivery has over a classroom, and it is the advantage most homeschooling families do not deliberately exploit.
• Point the parent's hour at the specific skill, not at reading together. Tutoring a named literacy skill outperformed listening to the child read by better than 2:1, and reading to the child produced nothing measurable in this literature.
• If a trained tutor in a scheduled slot is affordable, buy that instead. Teachers 0.50, paras 0.40, parents 0.23. Use the parent where the tutor cannot go — and see tutoring dosage for the recipe that makes tutoring work, because that recipe (3-5x/week, curriculum-aligned, a protected slot, the same person every time) is available to a homeschooling parent and is almost never what parent-tutoring studies actually tested.
• Spend the cheapest parental hour on monitoring, not teaching. Missed-assignment alerts bought +0.19-0.20 SD on GPA and +0.21 SD on maths; text nudges bought +0.11 SD against an active control. Neither needs the parent to know the subject.
• Do not select a tutor — including yourself — on credentials. A teaching certificate predicted nothing inside the largest homeschool dataset ever assembled, which is the same answer the archive's teacher-quality evidence gives about credentials generally.
• Do not buy the packaged full-service curriculum expecting it to carry the instruction. Enrolment in one made no difference inside Rudner's sample (F=0.24, n.s.), while structure — a defined lesson plan the parent actually executes — was worth 1.3 to 4.2 grade levels in Martin-Chang's comparison. The structure has to be practised, not purchased.
• Where the parent lacks the subject knowledge, outsource that subject rather than the child. This is the one place where the parent-education correlation plausibly has a causal core, and it is also the pattern most contemporary homeschooling families have already converged on.

Open questions

• Nobody has run the obvious trial: a parent tutoring their own child using the tutoring literature's known-good recipe — daily, protected, curriculum-aligned, with a coach checking fidelity. Every parent-tutoring study in the pooled evidence used the conditions known to fail.
• Does parental subject knowledge matter causally at the point of instruction? This is now the sharpest open question in the topic, because everything around it has been answered and it has not. The credential is a null; the parent's years of schooling is close to a causal null across three grade-B designs; who trains the parent is a measured null twice over. What nobody has varied is whether the parent can actually do the thing they are teaching. A trial that teaches the content to the parent before they teach it to the child would separate command-of-content from credential-and-class, and does not exist.
• Why does maths behave differently? Almost all parent-tutoring evidence is literacy; nearly all maths tutoring in the pooled evidence used paraprofessionals; the flagship parent-delivered maths trial is null on reanalysis; and the one clear maths finding is that an anxious parent helping with homework makes it worse. Home delivery's replicated weak spot is maths, and the parent-delivery literature has barely looked at it.
• What is the fidelity decay curve at home? The parent-implemented language trials measure fidelity and find it matters; no school-age instructional literature does. This is the single most likely explanation for the parent arm's underperformance and it is unmeasured.
• Why does the follow-up collapse happen, and can anything stop it? Two metas now measure it and agree it is severe (0.18 → 0.04; 0.50 → 0.16), with code-related skills surviving better than comprehension. Whether that is ordinary fadeout, the programme simply ending, or the parent reverting to unspecified activity is untested — and it is the difference between "re-dose" and "don't bother".
• Is there a floor below which a parent should not deliver instruction at all? Boulter's low-parent-education sample declined with more years of home education. That is one small longitudinal study carrying an uncomfortable and important question, and nothing has followed it up.