Perceived STEAM-integrated music education and students' innovative competence: parallel indirect associations through creative self-efficacy and context-specific learning engagement.

Innovative competence is an important educational outcome, but evidence from music-centered STEAM settings remains limited. This study examined whether students' perceptions of STEAM-integrated music education (SIM-E) were associated with innovative competence (IC) and whether creative self-efficacy (CSE) and context-specific learning engagement (LE) accounted for distinct portions of that association. A cross-sectional survey was completed by 427 middle school and university students in Shandong Province, China. Maximum-likelihood confirmatory factor analysis supported a four-factor measurement model, χ 2(428) = 489.87, CFI = 0.986, TLI = 0.985, RMSEA = 0.018, and SRMR = 0.038; the one-factor model fit substantially worse (CFI = 0.601, RMSEA = 0.099). HTMT values ranged from 0.319 to 0.533. Multigroup CFA further supported configural, metric, and scalar measurement invariance across middle-school and university participants (CFI = 0.975-0.977; absolute ΔCFI ≤ 0.002). In standardized composite-score regressions controlling for gender, age, prior music-learning experience, and prior STEAM experience, the adjusted total SIM-E-IC association was β = 0.331 (p < 0.001), and the residual direct association after CSE and LE was β = 0.194 (p < 0.001). With 5,000 bootstrap resamples, the specific indirect estimates through CSE (0.067, 95% CI [0.028, 0.111]) and LE (0.070, 95% CI [0.038, 0.106]) both excluded zero. Removing LE1, LE5, LE8, and LE9 yielded a reduced five-item engagement composite (LE-R); the LE-R indirect estimate remained 0.067 (95% CI [0.035, 0.103]), and the paired change from the full-LE estimate was 0.003 (95% CI [-0.010, 0.016]). An expanded-control sensitivity model that additionally adjusted for music major, digital-music-tool use, and school type also retained both specific indirect associations. Stage-specific bootstrap differences between middle-school and university estimates included zero. An exploratory ordered SIM-E → CSE → LE → IC product was also nonzero (0.040, 95% CI [0.024, 0.060]), but the cross-sectional design cannot establish temporal sequencing. Composite reliability was acceptable and the four-factor structure and discriminant validity were supported, whereas convergent-validity evidence remained mixed because AVE was below 0.50 for SIM-E, LE, and IC. The findings are therefore interpreted as robust parallel cross-sectional indirect associations rather than as causal or temporally ordered mediation.