Does teaching grammar improve children's writing?

Verdict

This is the sharpest divergence between popular practice and evidence in the archive. England made grammar a statutory, separately tested part of the primary curriculum in 2014 and expanded it beyond any previous version. Across the same period the evidence went the other way, and it did so in the designs that are hardest to argue with — large, independently evaluated randomised trials.

The verdict is negative rather than no-effect, and the distinction is deliberate. Against alternative writing instruction — which is the only comparison a teacher actually faces, because the hour is going to be spent on something — the meta-analytic estimate for traditional grammar teaching is d = −0.32 (Graham & Perin 2007), independently reproduced two decades earlier by Hillocks at −0.29. The modern, better-designed, contextualised versions do not reproduce that harm; they return zero. So the honest range is somewhere between costly and useless, and never beneficial for writing.

Three senses of "grammar" have to be kept apart or the answer becomes incoherent:

1. Grammar teaching as a means to better writing. This is what fails. Formal, isolated grammar (parts of speech, parsing, rules) fails with a negative sign. Contextualised grammar, embedded in writing lessons — the modern steelman — fails with a zero.
2. Sentence-level construction (combining, embedding, expanding sentences). Positive, and a genuinely different treatment, which is why it has its own topic. Note that Graham & Perin put grammar at −0.32 and sentence combining at +0.50 in the same table, and Andrews et al. ran two separate systematic reviews reaching opposite conclusions. Treating them as one thing is the single most common error in this literature.
3. Transcription and mechanics — spelling, punctuation, handwriting. These are separate teachable skills with their own evidence and their own topics (spelling, handwriting & transcription). "Grammar" in ordinary speech often means these, and they behave completely differently.

What the evidence shows

The strongest evidence is the most deflationary, and it gets more deflationary as the design improves. The developer-led trial found +0.21. The EEF efficacy trial of the same programme found +0.10, not significant. The EEF effectiveness trial — longer dose, more training time, 155 schools, independent evaluators — found −0.02, plus a negative effect on the national GPS test. That is the classic efficacy-to-effectiveness decay curve, and here it decays through zero.

The Englicious trial closes the last escape route. Every previous null could be waved away as "wrong age, wrong dose, wrong kind of grammar." Wyse et al. ran the first robust RCT on six- and seven-year-olds, using a modern-linguistics-informed programme in which every lesson ended with children applying the grammar to their own writing. Narrative writing moved d = 0.04. Their own conclusion: "The study finding does not offer support for grammar teaching to improve writing."

Teacher satisfaction is uninformative and the EEF trial proves it. More than 90% of intervention teachers found the programme, training and materials useful. The effect was −0.02.

The one thing the EEF trial did not find is the thing grammar teaching is supposed to be for. Not only did writing not improve — the KS2 grammar, punctuation and spelling score went slightly down, and teachers' own grammar knowledge on a quiz did not improve either.

Hereditarian-lens assessment

Risk: low, and this topic is unusual in that the archive's premise barely has to do any work.

The verdict does not rest on correlational evidence at all. It rests on three independently evaluated randomised controlled trials — 155 schools, 70 schools, 53 schools — plus four meta-analyses of experimental and quasi-experimental studies. Treated and untreated children in a school-randomised trial have the same expected distribution of language-relevant alleles. There is no selection story that produces a null.

Where heredity does bite is on the one positive result. Jones, Myhill and Bailey found that contextualised grammar "benefit[ed] able writers more than weaker writers." That is a within-arm subgroup contrast on a substantially heritable trait. Olson et al. 2013 finds substantial genetic influence on writing samples and handwriting copy in 540 twins aged 8–18, with shared environment not significant — the household does not explain who writes well. So "able writers" is largely a genetic category, and "the intervention helped the able writers" is an association between a heritable trait and a response, not a demonstrated aptitude-treatment interaction. The EEF effectiveness trial tested the same moderation at scale with proper randomisation and found the opposite sign: no benefit at or above the median, a small negative effect below it.

One more consequence of the premise worth stating, because it cuts against a common defence of grammar teaching: if you did find that explicit grammar helped the strongest writers most, that would be an argument for it as an enrichment for children who already write well — not as a remedy for children who do not, which is how it is nearly always sold.

Boundaries & what critics say

• The steelman is Jones, Myhill & Bailey (2013), and it deserves to be stated at full strength. Their argument is that every earlier meta-analysis tested isolated grammar, and that grammar taught as a meaning-making resource embedded in writing instruction is simply a different intervention. Their trial found e = 0.21, p < 0.001, in 744 Year 8 pupils. It is the best-theorised positive result in the field. Three things weigh against reading it as settled: Wyse & Torgerson (2017) show it used no intention-to-treat analysis and analysed at pupil level despite cluster-level randomisation and delivery, both of which inflate precision; the study was developer-led; and two subsequent independent EEF trials of the programme it produced found 0.10 (ns) and −0.02.
• Rogers & Graham (2008) is the honest counter-citation and should be given. A meta-analysis of 88 single-case design studies identified nine supported writing treatments — including "teaching grammar and usage" and "teaching sentence construction skills." It is the only synthesis pointing that way. It is also a corpus with no control groups anywhere in it, in which participants serve as their own baselines, and in which "grammar and usage" typically means targeted correction of one error class in one struggling writer, measured on that same error class. That is mechanics remediation, not a grammar curriculum.
• The small-group result is a trap, and it is the trap most relevant to a reader with three children. Torgerson et al. (2014) found ES = 0.24 for children who got Grammar for Writing in small groups on top of whole-class teaching, versus 0.10 for whole-class alone. That looks like "grammar works in small groups." The evaluators say otherwise on p. 33: the small-group children were "not much higher than those taught in larger groups in either the control or intervention conditions", so "this difference could have been as a result of teaching children in small groups per se rather than as a result of the programme." Small groups are doing the work; the content is along for the ride.
• Nobody claims grammar is unteachable. It is perfectly teachable as a body of knowledge. The claim being rejected is a transfer claim — knowing the names of word classes does not make the sentences better. This is the same structure as comprehension strategies and motor competence transfer: the trained thing improves, the thing it was supposed to cause does not.
• The evidence base for the negative number is old and heterogeneous. Graham & Perin's grammar category is k = 11 with a significant heterogeneity test and an unweighted SD of 0.61; many of the underlying studies are decades old quasi-experiments. Treat −0.32 as "the older literature says this actively cost something" and the modern RCT nulls as the better estimate of what a contemporary programme does.
• Writing quality is a subjective rating almost everywhere in this table. That inflates positive findings. It is worth noting that it should inflate grammar findings too, and grammar is still the only negative entry.

Practical guidance

• Do not buy a grammar programme to improve writing. Not at 6, not at 10, not at 13. Three randomised trials and four meta-analyses is more converging evidence than almost anything else in this archive has.
• If you have already bought one, the opportunity cost is the argument, not the harm. Six weeks of daily hour-long grammar lessons displaced six weeks of something else, and in the trial that measured it the displaced something else was worth about a month of GPS progress.
• Teach sentence-level construction instead — combining, expanding and embedding sentences — which is a different treatment with the opposite sign. See sentence combining, including its honest age boundary.
• Below about age 7, do not expect grammar work to do anything. The Englicious trial is the only robust evidence at this age and it found d = 0.04. The developmental note in that report is the useful one: the ability to combine sentences using conventional grammar arrives around age seven for most pupils. Before that, the binding constraints are transcription and oral language, not grammatical knowledge.
• Distrust your own satisfaction. 90% of the teachers in the null trial liked the programme.
• Teach grammatical terminology if a test requires it, and be honest that that is why. It is a small, learnable body of knowledge. It is not a writing intervention, and even the trial designed to improve GPS scores did not improve them.

Open questions

• Confidence is capped at medium by process, not by the evidence. Three grade-B randomised trials agree, which would ordinarily clear the bar for high; status: surveyed means the adversarial pass METHODOLOGY requires has not been run against this verdict.
• Nobody has tested whether the sentence-generation signal in the Englicious trial is real. d = 0.14 was not significant, but it was the grammar-proximal measure and the authors think it reflects the manipulation of words and phrases rather than the naming of them. That is a testable hypothesis and it points straight at sentence combining.
• No trial has isolated group size from content. Torgerson 2014's 0.24 is the only randomised small-group contrast in the topic and it cannot separate the two. Given that the archive's motivating question is about three children, this is a real gap.
• Whether contextualised grammar works for secondary-age able writers specifically is unresolved. Jones, Myhill & Bailey's Year 8 result has never been independently replicated at that age with intention-to-treat analysis at the level of randomisation.