Fostering scientific reasoning in education – meta-analytic evidence from intervention studies

~30 interventions from experimental and quasi-experimental studies (conference version); primary through secondary education

Primary and secondary education; the field's own synthesis of everything that calls itself a scientific-reasoning intervention.

The field's headline number for "scientific reasoning can be taught". Aggregates experimental and quasi-experimental studies whose outcome is a scientific-reasoning measure, which in this literature means a researcher-designed instrument built around the intervention's own target construct (control of variables, hypothesis testing, evidence evaluation). Moderator analysis reports that constructive activities out-performed interactive ones. CAUTION ON THE NUMBER: g = 0.80 over ~30 interventions is the figure from the ISLS 2015 conference version of the same meta-analysis (Engelmann & Fischer, "Fostering Scientific Reasoning: A Meta-analysis on Intervention Studies"); the 2016 journal version was paywalled and its headline estimate was not independently confirmed here. Treat the value as approximate until someone reads the journal article. No transfer to another school subject, to a general reasoning test, or to any independent standardised achievement measure is aggregated — the meta-analysis asks whether the trained construct moves, and answers yes.

Scientific reasoning is trainable, at a large apparent magnitude (g ≈ 0.80), and the success of an intervention depends on its content, pedagogy and — decisively — its assessment. That last moderator is the finding a founder should carry away: the effect is measured on instruments aligned to what was taught. Read alongside the archive's 2x measure-inflation correction and the fact that no meta-analysis in this area pools an independent standardised outcome, a g of 0.80 on aligned measures is consistent with a near-zero effect on anything a school or examination board measures — which is what the one well-powered randomised effectiveness trial in this topic (EEF 2016) actually found.

Medium: pools experimental and quasi-experimental designs, so selection into treatment is present in part of the base. Not the main threat here — measure alignment is.

The trainability finding is consistent across this literature (Chen & Klahr 1999; Klahr & Nigam 2004; Zohar & Nemet 2002; Antonio & Prudente 2024). What has never replicated is the step from trained-construct gains to gains on independent measures of anything else.