Evidence-based writing instruction: what works.

The neglected R

Of the three Rs, writing gets the least science and the most folklore. Reading has meta-analyses that changed national policy. Mathematics has decades of cognitive research. Writing, in most classrooms, is taught the way it was taught to the teacher: assign a piece, mark the errors, drill some grammar, repeat. National assessments keep returning the same verdict — most students write below the level their futures will demand.

The folklore was never necessary. Controlled experiments on writing instruction have accumulated since the 1960s, and they have been pooled twice at scale: once for adolescents (Graham & Perin, 2007), once for the elementary grades (Graham, McKeown, Kiuhara & Harris, 2012). The result is something writing pedagogy rarely gets: a ranked list of teachable practices, each with a measured average effect on the quality of what students write.

This article walks the list — what tops it, what surprises, and why the most familiar staple of language teaching comes out below zero. Then it turns the rankings into a working plan for schools.

What Writing Next measured

In 2007, Steve Graham and Dolores Perin published Writing Next, a Carnegie-commissioned meta-analysis of writing instruction for grades 4 to 12 (Graham & Perin, 2007). They gathered every true or quasi-experiment they could find in which one group of students received an identifiable instructional treatment, another did not, and writing quality was measured at the end. A companion analysis in a peer-reviewed journal reports the method in full, drawing on well over a hundred studies (Graham & Perin, 2007).

The output was eleven instructional elements, each with an average effect size — the difference treatment made, in standard deviations, on judged writing quality. An effect near 0.20 is small; 0.50 is moderate; 0.80 is large for a school intervention. Because every element faced the same yardstick, the list reads as a league table. Two cautions before the table. The elements are not rivals — several combine naturally. And effects on judged quality are the field’s standard outcome, but they carry the noise of human rating (Graham & Perin, 2007).

It helps to feel what the numbers mean in a classroom. An effect of 0.82 implies that the average student in the treatment group ends up writing better than roughly four out of five comparison students. An effect of 0.50 moves the average student past about two-thirds of them. These are not laboratory curiosities. They are differences a teacher can see in a stack of marking — and, at the top of the table, differences most schools have never deliberately bought.

The league table

Top of the table, at an average effect of roughly 0.82: strategy instruction — explicitly teaching students strategies for planning, drafting and revising, modeling each one, then fading support until students run the process alone (Graham & Perin, 2007). Tied with it, summarization at roughly 0.82: teaching students to compress a text into its essentials, a compact exercise in choosing and structuring content.

Collaborative writing follows at ≈0.75 — structured arrangements where students plan, draft and revise together rather than alone. Specific product goals land at ≈0.70: telling students precisely what the finished piece must accomplish, not just its topic. Word processing earns ≈0.55, sentence combining ≈0.50, and further down sit prewriting, inquiry and the process-writing approach around ≈0.32, with study of models at ≈0.25 (Graham & Perin, 2007).

Notice what the top of the table shares. Strategy instruction, summarization, collaboration and product goals all make the invisible work of writing visible — planning, selecting, structuring, revising against a target. None of them is about correctness. The elements that police surface features cluster at the bottom. And one of them falls off the table entirely.

The grammar result

The most useful number in Writing Next is its most uncomfortable one. Across the studies in which traditional grammar instruction — parts of speech, sentence parsing, usage drill — served as a treatment, the average effect on writing quality was roughly −0.32 (Graham & Perin, 2007). Not zero. Negative. Students receiving grammar-focused instruction wrote measurably worse, on average, than comparison students spending the same time otherwise.

The finding was not new; it was a replication. George Hillocks’ landmark 1986 review of two decades of composition experiments had reached the same figure. Traditional grammar teaching scored near −0.3 — the worst-performing focus of instruction he examined — while sentence combining sat comfortably positive at ≈+0.35 (Hillocks, 1986). Twenty years later, a systematic review at York combed the trial evidence again and found no evidence that teaching formal grammar improves the quality or accuracy of children’s writing (Andrews et al., 2006). Three sweeps of the literature, one answer.

Why does a result this consistent keep losing to tradition? Partly habit: grammar drill is easy to set, easy to mark and visible to parents. Partly the exams: where tests reward error-spotting, teaching drifts toward it. And partly a confusion between two goals. Grammar as knowledge about language is a legitimate subject in its own right. The trials only measure its value as a route to better writing — and as a route, it points backwards (Andrews et al., 2006).

The constructive replacement is sentence combining: students take kernel sentences — The dog barked. The dog was old. — and fuse them into better ones, in the context of their own writing. That is grammar as a production skill rather than an identification skill. In a randomized study with fourth graders, peer-assisted sentence-combining practice beat grammar instruction on sentence-construction skill, with gains visible in story quality — for stronger and weaker writers alike (Saddler & Graham, 2005).

Writing in the elementary grades

Everything so far concerns grades 4 and up. In 2012, Graham, McKeown, Kiuhara and Harris pooled the true experiments for the elementary years, and the table replicated with the volume turned up (Graham, McKeown, Kiuhara & Harris, 2012). Strategy instruction again topped the list, with an average effect near ≈1.0 — and larger still when the strategies came wrapped in self-regulation: goal setting, self-monitoring, self-talk, the package known as SRSD, developed by Harris and Graham themselves.

The supporting cast repeated too, at approximate effects worth reading as a set: peer assistance around ≈0.9, specific product goals ≈0.76, prewriting activities ≈0.54, word processing ≈0.47, and simply adding writing time ≈0.30 (Graham, McKeown, Kiuhara & Harris, 2012). Even the humblest entry carries a lesson: many elementary classrooms allocate writing so little time that merely scheduling more of it produces a measurable gain.

For school leaders the elementary table removes an excuse. The strongest practices are not grade-locked. A seven-year-old can be taught a planning strategy, a revision checklist, a way to talk herself through a draft — and the measured payoff is, if anything, larger than for teenagers (Graham, McKeown, Kiuhara & Harris, 2012).

Why strategy instruction wins

The league table has a theory underneath it. Ronald Kellogg’s account of how writing skill develops treats composition as one of the most cognitively demanding things schools ask of anyone. Idea generation, word choice, sentence construction, spelling and handwriting all compete for the same limited working memory at the same moment (Kellogg, 2008). Experts survive the load because large parts of the process have become automatic, freeing attention for meaning. Novices drown in it.

Read the winning elements against that account and they stop looking like a grab bag. Strategy instruction hands the novice an expert’s routine — plan first, using these steps; revise against these questions — so the process no longer has to be improvised under load. Product goals shrink the target. Collaboration distributes the load across heads. Sentence combining and transcription practice automate the sentence-level machinery, which is why fluent handwriting and typing matter well beyond neatness (Kellogg, 2008). Every high-ranking element is a load-management device.

It also explains the self-regulation bonus in the elementary data: writing is long, frustrating work, and strategies survive only if the writer can monitor herself through them (Graham, McKeown, Kiuhara & Harris, 2012). Teaching the strategy without the self-management is handing over a map without the compass.

Writing feeds reading

One more pooled result belongs in the case for taking writing seriously. In 2011, Graham and Michael Hebert assembled the experiments on writing’s effect on reading. Writing about a text improved comprehension of it by roughly a third of a standard deviation on average — outperforming rereading it or discussing it (Graham & Hebert, 2011). Teaching writing, and simply increasing how much students write, also strengthened reading skill more broadly.

For curriculum planners this dissolves the zero-sum framing in which writing time competes with reading time. Summarizing, answering in writing, note-making — writing is how a reader externalizes and tests understanding. The two Rs are one system, and the league table above is also, quietly, reading instruction (Graham & Hebert, 2011).

What the evidence doesn’t show

The writing literature is unusually consistent, but its findings are narrower than the slogans they spawned. The honest limits:

• That grammar knowledge is worthless. The negative effect attaches to decontextualized drill displacing writing practice; grammar exercised inside composition — sentence combining — is one of the table’s solid positives (Hillocks, 1986) (Saddler & Graham, 2005).
• A ready-made curriculum. The meta-analyses rank elements, not sequences; how much strategy instruction, combined with what, in which genre and grade, remains professional design work (Graham & Perin, 2007).
• Durability and transfer. Most trials measure writing soon after treatment, in genres close to those taught; evidence on maintenance over years, and transfer to unfamiliar genres, is far thinner (Graham & Perin, 2007).
• An objective outcome. Effects are on judged writing quality — human ratings with real reliability limits; the numbers inherit that noise (Graham & Perin, 2007).
• That devices improve writing. Word processing’s ≈0.5 reflects drafting and revising fluidity, not hardware; a laptop scheme without instructional change buys the tool effect at most (Graham, McKeown, Kiuhara & Harris, 2012).
• AI-era evidence. Every pooled study predates generative AI; how drafting with a model reshapes the practice that builds writing skill is, so far, unmeasured.

Where the evidence stops

1. 1That grammar knowledge is worthless
2. 2A ready-made curriculum
3. 3Durability and transfer
4. 4An objective outcome
5. 5That devices improve writing
6. 6AI-era evidence

© 2026 FUTURE PROOF™

Writing instruction by the evidence

Turned into practice, the league table becomes six moves a school can make this term.

Teach the process explicitly. Name a planning strategy, model it aloud, practise it with support, fade the support — for every major genre. This is the table’s top row, in both age bands (Graham & Perin, 2007) (Graham, McKeown, Kiuhara & Harris, 2012). Add the self-regulation wrapper: goals, self-checks, self-talk.

Swap drill hours for sentence combining. Take the timetable slots that grammar drill occupies and refill them with combining, expanding and imitating sentences inside students’ own drafts (Hillocks, 1986) (Saddler & Graham, 2005). Same grammatical content, opposite sign.

Give every assignment a product goal. Not “write about the trip” but “persuade a reader who disagrees, with three reasons and a counterargument answered”. Specific targets earn ≈0.7 on their own (Graham & Perin, 2007).

Schedule writing, and some of it together. Volume alone helps in the elementary years, collaboration helps everywhere — structured pairs planning, drafting and revising jointly (Graham, McKeown, Kiuhara & Harris, 2012).

Automate the machinery early. Fluent transcription — handwriting and typing — frees working memory for composing; the load account predicts, and the data confirm, that sentence-level automaticity is a prerequisite, not a nicety (Kellogg, 2008).

Write across the curriculum. Summaries and written responses in science and history lift reading comprehension while building writing mileage — one intervention, two Rs (Graham & Hebert, 2011).