The Challenge of Improving Children's Writing Ability: A Randomized Evaluation of "Writing Wings"

39 schools in 21 states, 152 classroom teachers, about 3,000 pupils randomly assigned; impact sample smaller after attrition

Grade 3-5 pupils (roughly ages 8-11) in high-poverty US elementary schools

A cluster randomised evaluation of Writing Wings, the Success for All Foundation's writing curriculum, run over two annual cohorts. Two classrooms per grade per school were randomised to the programme or to business as usual, so teacher-level randomisation is nested inside schools. The reason this study matters to a conventions topic out of all proportion to its fame is its outcome structure: a researcher-administered fall and spring writing prompt scored 1-5 on FOUR SEPARATE DIMENSIONS, one of which is Conventions, defined in the report as "proper use of grammar, paragraphing, punctuation, and spelling". It is, as far as this archive has been able to establish, the ONLY randomised evaluation of a whole writing curriculum that scores conventions as its own dimension and reports it. The measure is a rubric dimension applied to children's own composition, not an independent standardised conventions test - which makes it a measure of conventions USE rather than conventions knowledge.

A precisely estimated zero on every dimension, conventions included. Conventions: 2.34 treatment vs 2.32 control, difference 0.021, effect size 0.02; the HLM treatment coefficient was gamma_010 = 0.006, not significant. By grade the conventions difference was -0.026 (Grade 3), +0.065 (Grade 4) and +0.016 (Grade 5), none significant. Total writing score 2.46 vs 2.44, ES 0.03; word choice 0.06; content and audience 0.01; organisation 0.01. Teacher ratings of pupils' writing likewise showed no differences. The covariate structure is a reminder of what actually predicts conventions in this sample: grade level (Grade 5 +0.177), sex (male -0.162) and ethnicity coefficients are all several times larger than anything the curriculum did.

Low. Teachers and classrooms were randomly assigned within schools, so treated and control pupils have the same expected distribution of ability-relevant alleles. The demographic coefficients in the model are descriptive associations in a non-randomised dimension and carry no causal content.