Recalibrating impact in STEM education: Empirical benchmarks for interpreting intervention effects

Effect sizes are central to evaluating intervention impact in STEM education, yet they are commonly interpreted using generic benchmarks such as Cohen's thresholds, which may misrepresent practical significance in this heterogeneous field. To address this problem, the present study derived empirical effect size benchmarks for interpreting intervention effects in STEM education. We conducted an umbrella review of meta-analyses on STEM education interventions and synthesized 95 meta-analyses comprising 5,956 effect sizes from 3,610 primary-study entries. Benchmarks were estimated from the distribution of absolute Hedges' g values overall and across four analytic layers: STEM field, intervention category, outcome domain, and valence. Publication-bias-adjusted benchmarks were estimated using limit meta-analysis, and robustness was examined via precision weighting and a one-per-primary-study specification. Overall raw quartiles were 0.21, 0.48, and 0.86 at the 25 th , 50 th , and 75 th percentiles, respectively, but these declined to 0.14, 0.32, and 0.59 after publication-bias adjustment. Substantial heterogeneity emerged across STEM fields and intervention categories: science and cognitive, metacognitive, and strategic learning supports showed the largest benchmarks, whereas mathematics and assessment/contextual interventions showed smaller benchmarks. Academic and nonacademic outcomes were comparatively similar, whereas desired outcomes showed markedly larger benchmarks than undesired outcomes. Robustness analyses indicated that subgroup ordering remained largely stable despite more conservative effect size interpretation benchmark magnitudes under alternative specifications. These findings show that effect size interpretation in STEM education should rely on empirically derived, context-sensitive benchmarks rather than universal conventions.