The impact of an integrated approach to science and literacy in elementary school classrooms

94 grade-4 teachers/classrooms in 16 districts in one southern US state; ~2,050 students analysed (1,027 treatment / 1,026 comparison); writing subsample n = 467

US grade-4 classrooms, urban/suburban/rural, ~55% free-or-reduced-price lunch; an 8-week (40-session) unit on light and energy.

Cluster RCT: teachers were randomly assigned to teach the integrated Seeds of Science/Roots of Reading light unit or a content-comparable science-only unit from their own district materials plus their regular literacy instruction. Three-level HLM (students in classrooms in schools). Attrition ~5%. THE INDEPENDENCE PROBLEM IS TOTAL: Barber, Dorph and Pearson are at UC Berkeley's Lawrence Hall of Science, which developed Seeds/Roots, and Cervetti was on the development team — the developers evaluated their own curriculum. ALL FOUR OUTCOME MEASURES ARE RESEARCHER-DESIGNED and were built at Lawrence Hall of Science 'in partnership with the curriculum research and development team', with items developed alongside the treatment curriculum during its field test. The team did make a real concession to fairness: the science-understanding analysis uses only a 23-item subset that external science experts judged fair to both conditions, and the vocabulary words all appeared in the state standards guiding the comparison teachers. But no independent or standardized measure was administered. Slavin et al. (2014) identify a second confound the paper does not resolve: treatment teachers taught 3.66 hours of science per week vs 3.03 in the comparison group (ES = +0.53, p<.05), so the integration effect is entangled with DOSE.

The pattern is exactly the archive's default: big on the aligned measures, nothing on the transfer measure. Science understanding ES = 0.65, science writing (7-dimension rubric, multivariate) ES = 0.40, science vocabulary ES = 0.22, science-word count in writing ES = 0.80 — all on instruments built alongside the curriculum. Reading comprehension: ES = +0.09, NOT SIGNIFICANT — and note that this was the developers' OWN researcher-designed comprehension measure, not a standardized one, so the far-transfer claim for science-literacy integration fails on its home turf. The honest summary for a founder: teaching a coherent, knowledge-rich, literacy-rich science unit reliably teaches THAT science content and the vocabulary in it, and does not measurably improve general reading comprehension within 8 weeks. Treat 0.65 as an upper bound inflated by both developer-led evaluation and aligned measurement; the archive's writing tranche found a ~5x aligned-to-standardized gap and the honest expected standardized value here is a fraction of 0.65.

LOW for the internal comparison — randomization at teacher level defeats passive gene-environment correlation. The threats here are measure alignment, developer involvement and the unequal science-time dose, not heredity.

Mixed. The independent best-evidence synthesis of elementary science programmes (Slavin, Lake, Hanley & Thurston 2014, JRST 10.1002/tea.21139) includes this study and treats it as one of the better-designed science-literacy integration trials while flagging the dose confound. The broader integration literature replicates the PATTERN but not the magnitude: Hwang, Cabell & Joyner (2022) find g = 0.54 on researcher-developed comprehension measures but only 0.25 on standardized ones, and no significant effect on standardized vocabulary at all. Vitale & Romance's Science IDEAS reports larger effects on genuinely standardized tests (MAT-Science +0.66 to +0.90, ITBS-Reading +0.11 to +0.40) but is developer-led and non-randomized.