Teaching children to write

The program (what a school should actually do)

Writing is not reading. Reading is close to a solved problem at the level of method, with large effects, converging designs and a clear sequence. Writing has fourteen studies in the entire Grade 2–12 range that meet an independent-measure standard, an average effect of +0.18 across them, and a curriculum in England that was rewritten in 2014 around the one practice the evidence most consistently rejects.

So the honest programme is short, and most of it is about what not to do.

Ages 5–7 — build transcription, not grammar.

• Teach letter formation explicitly: model the stroke sequence, use numbered arrows showing stroke order and direction, then have the child write the letter from memory. That specific combination beat four other handwriting methods in the one trial that compared them head to head. Roughly 15–20 minutes, twice a week, for a term, to the children who need it — the randomised trials that established this ran in groups of three. Then have them write something immediately afterwards.
• Skip fine-motor and perceptual-motor handwriting programmes. Pure motor instruction returns ES ≈ 0.10 for legibility and −0.07 for fluency. It is the clearest waste of money in the subject. (evidence: handwriting & transcription)
• Teach spelling explicitly and systematically (ES 0.54 against not teaching it; 0.70 for more explicit over less explicit) — and justify it on spelling and reading, which it demonstrably improves, not on composition, which it does not. (evidence: spelling)
• Do not teach grammar to improve writing at this age. The only robust trial ever run on six- and seven-year-olds found d = 0.04 on narrative writing. The ability to combine sentences using conventional grammar arrives around age seven; before that the binding constraints are transcription and oral language.

Ages 8–12 — sentence-level work plus an explicit composing routine.

• Sentence combining — give two or three short sentences, have children combine them, then apply the move inside their own draft. It is free, needs no programme, and it is the one member of the grammar family that meta-analyses score positively (+0.50 against grammar's −0.32 in the same table). Note the age floor: the evidence base runs from about Grade 4, not from age 5. (evidence: sentence combining)
• Teach an explicit, named, memorised routine for planning, drafting and revising, model it, and fade the support. This is the highest-value item in the writing curriculum and one of the only writing practices that survives an independent-measure standard at all. Budget +0.15 to +0.20, not the +0.82 to +1.14 the meta-analyses advertise. (evidence: writing strategy instruction)
• Add morphology early — prefixes, suffixes, roots — folded into the literacy block rather than run as its own unit. It helps younger students as much as older ones and helps weaker readers most.

Throughout — two conditions matter more than which programme you buy.

• Train the teachers properly. The process writing approach is worth 0.46 where teachers received professional development and 0.03 where they did not (−0.05 in Grades 7–12). The same programme returned ~+0.74 when its developers trained the teachers and +0.11 when newly recruited trainers did. Training is not an implementation detail; it is most of the effect.
• Count what writing time displaces. In the only two-year trial, pupils doing an intensive writing programme made significantly less progress in reading, spelling and maths. Writing hours come out of something.

What to refuse to spend money or time on

• Grammar programmes sold as writing interventions — the best-evidenced failure in this subject. Grammar is the only negative entry in the largest adolescent writing meta-analysis (−0.32), the only one of thirteen elementary treatments to miss significance, and a 155-school randomised trial returned −0.02 on writing plus a small negative effect on the national grammar test.
• Fine-motor / perceptual-motor handwriting programmes — 0.10 for legibility, −0.07 for fluency.
• Spelling programmes marketed as improving composition — the purpose-built meta-analysis explicitly failed to find that transfer.
• Any writing programme quoting an effect size above ~0.5 without saying who scored the writing.
• Switching a struggling young writer to a keyboard as a workaround — handwriting and keyboarding speed correlate at r ≈ .43, and the early-writing review mostly favours handwriting.
• Teacher satisfaction as evidence. In the null EEF grammar trial, over 90% of intervention teachers found the programme, training and materials useful. The effect was −0.02.

The measurement problem, which is this subject's defining feature

Every other subject in this archive has a measure-type problem. Writing has it worse than any of them, and a founder who does not internalise this will buy the wrong thing.

Writing quality is a human judgement of the overall merit of a composition, made in almost every study by someone who knows what was taught. That is the most inflatable outcome class in education. The consequences are visible whenever the same literature is re-filtered:

METHODOLOGY's nominal correction is 2×. In writing it is closer to 5×. Goal setting appears in one respected meta-analysis at d = 2.03; nothing in education moves an independently measured outcome two standard deviations, and that number's presence at the top of a real table is the best available warning about the rest of it.

The hereditarian bottom line for a founder

In 540 twins aged 8–18, writing samples and handwriting copy show substantial heritability and shared environment is not significant — the household does not explain who writes well, any more than it explains who reads well or who has aerobic capacity. Three consequences a school builder should hold onto:

1. Composition and transcription are partly genetically distinct. Genetic influences across writing measures are correlated — most strongly between the two speeded ones, writing fluency and handwriting copy — but there are also significant independent genetic influences separating handwriting copy from writing samples. So automating transcription is worth doing on its own terms; it is not a hidden route to better composition, and the newest K-3 meta-analysis finding no transcription effect on any composition dimension is what that partial independence predicts.
2. Every "it helped the able writers most" finding is an association, not an aptitude-treatment interaction. The one positive grammar trial found contextualised grammar benefited able writers more than weaker ones. "Able writer" is largely a heritable category, so that is a correlation between a heritable trait and a response. When the same moderation was tested at scale with proper randomisation, the sign reversed: no benefit at or above the median, a small negative effect below it.
3. Aim at floor-raising and expect heritability to rise, not fall. Nobody spells, forms letters or plans an essay without being taught. Good instruction gets every child over those thresholds and thereby increases the share of remaining variance that is genetic. Per METHODOLOGY that is evidence the teaching worked.

Nothing in this subject bears on g.

Confidence

Open questions a founder should watch

• The whole subject rests on fourteen independently measured studies across eleven year groups. That is the most important fact about it, and it is a fact about the field rather than about any finding.
• Almost nothing here reaches ages 5–7. Grammar has one trial at that age (null). Sentence combining has none. Strategy instruction has a K-3 meta-analysis. Only handwriting and spelling have direct randomised evidence in the first two years of school — and the handwriting trials are from 1997 and 2000 with n = 144 and n = 38.
• No modern, adequately powered, independently evaluated RCT of sentence combining exists, which is awkward given it is the practice this archive ends up recommending.
• Nothing in this subject measures durability beyond about six months, and only one study reaches that far.
• The transcription-to-composition contradiction is unresolved: two randomised first-grade trials say it transfers, the largest K-3 meta-analysis says it does not, and the reconciling hypothesis (it transfers only where transcription is the binding constraint) has never been tested.