Cognitive science Teacher Choices trial: Using examples to teach grammar to Year 7. Evaluation report

55 secondary schools; 228 teacher-class units randomised; 6,297 pupils analysed on writing and 6,642 on the grammar assessment

Year 7 pupils (aged 11-12) in English secondary schools

Three-armed cluster RCT randomised at teacher-class-unit level, independently evaluated by NFER for the EEF, October 2025. IMPORTANT SCOPE LIMIT: this is a "Teacher Choices" trial, so all three arms teach grammar and there is NO business-as-usual arm - it tests HOW to use examples (systematic worked, responsive worked, non-worked), not WHETHER to teach grammar. Around 20 sessions of 15 minutes over 10 weeks. The primary outcome was a bespoke Text-type Specific Writing Assessment scored on three strands, one of which is PUNCTUATION (the others being sentence structure/text organisation and composition and effect). The secondary outcome, the Noun Phrase Grammar Assessment (NPGA), is a 30-item closed-response grammar test built by NFER largely from released KS2 GPS past-paper items - the closest thing in the UK trial literature to a standardised conventions subtest. Well powered (MDES 0.25 achieved) but 29% missing on the primary outcome, evenly spread across arms. Its most valuable content for the conventions question is not the impact estimate but Table 19, an incidental large-n validity table nobody set out to produce.

Nulls throughout. Writing composition: adjusted means 18.17, 17.92 and 17.60 out of 40 with overlapping CIs; no evidence of difference between approaches. Grammar knowledge (NPGA): likelihood-ratio chi-square = 0.18, df = 2, p = 0.912, n = 6,642, ICC 0.04; adjusted means 17.61, 17.64 and 17.56 - so flat that no effect size was computed. THE VALIDITY RESULT IS THE IMPORTANT ONE: in Table 19, pupils' KS2 GPS score correlated r = 0.79 with the bespoke grammar assessment (n = 6,642, p < .001) but only r = 0.56 with the independently marked writing assessment (n = 6,297, p < .001). A national conventions test explains roughly 62% of the variance in another decontextualised grammar test and roughly 31% of the variance in what children actually write, in the same 6,000-pupil sample. The process evaluation supplies the mechanism in the teachers' own words: pupils could use a taught grammar pattern "when this was highly scaffolded", but "teachers perceived that pupils rarely transferred use of the taught grammar patterns into more general writing composition tasks", and transfer "into writing tasks where use of the pattern did not immediately follow modelling, or was not explicitly prompted, was much rarer". The report frames the whole problem as "the transfer of declarative knowledge into procedural application in original writing".

Low for the impact estimates - randomisation at teacher-class level with an independent evaluator, and baseline balance on KS2 GPS confirmed (chi-square 2.43, df 2, p = 0.296, n = 8,446). The Table 19 correlations are observational and confound general ability, which inflates BOTH coefficients; the informative quantity is the 0.79 vs 0.56 gap within the same pupils, since whatever general ability contributes is common to both.