Construction of a hierarchical linear regression prediction model for classical Chinese poetry writing ability development
چکیده
Although classical poetry is central to Chinese language education, how poetry learning is linked to writing development has rarely been characterized empirically. This one-year longitudinal study followed 1,920 students in Grades 3–6 from 48 classes in 12 primary schools across four waves and applied a four-level growth model (measurement occasions nested within students, students within classes, and classes within schools). The cognitive load components were indexed provisionally by NASA-TLX subscales, and results that depend on their distinction are interpreted as preliminary. Eight confirmatory tests were controlled with the Holm procedure and all remaining analyses were reported as exploratory with Benjamini-Hochberg control; a random-intercept cross-lagged panel model examined within-person lagged associations, which inform rather than establish the ordering of the modeled associations; and the 12-school highest level was examined with leave-one-school-out refits, a wild cluster bootstrap, prior sensitivity analyses, a school fixed-effects model, and a calibrated Monte Carlo simulation. Intrinsic (β = −0.41) and extraneous (β = −0.49) cognitive load were negatively and germane cognitive load (β = 0.45) positively associated with writing ability (Holm-adjusted p < .001), and schema construction, imagery processing, and emotional resonance carried statistical indirect associations totaling 0.587 (76.2% of the total association). The interaction of teaching method with cognitive load ( γ = 0.47, Holm-adjusted p = .011) was retained under all small-cluster sensitivity analyses, whereas the interactions with grade and class environment were exploratory and the school-level variance component was imprecisely estimated, and writing ability followed a mildly decelerating quadratic trajectory ( π 2 = −0.17 points per wave², Holm-adjusted p < .001). With imputation, centering, model reselection, refitting, and BLUP computation confined to the training folds, prediction for new students in the sampled classes and schools reached R 2 = 0.847 and RMSE = 4.28, meeting the preset RMSE target, whereas three-band accuracy (80.3%) fell short of the preset 85% target, and prediction for unseen classes and schools reached R 2 = 0.768 and 0.656. The findings characterize the multilevel associations between classical poetry learning and writing development and support instructional recommendations restricted to the confirmatory results.