Zhou Bairen

Introduction This study treats adolescent athletes' self-efficacy as a dynamic process rather than a fixed quantity, asking how its fluctuations arise in online social settings and how that volatility shapes training engagement. Methods We recruited a stratified sample of 320 youth athletes across sport discipline, competitive tier, and sex, and combined ecological momentary assessment with weekly follow-up over 8 consecutive weeks, using instruments released under open licenses and open-source software throughout. Two dispersion indices, the intraindividual standard deviation and the root mean square of successive differences, summarized fluctuation at the person level, while absolute occasion-to-occasion change indexed momentary instability at the event level. We fitted multilevel linear models, a two-level dynamic structural equation model, latent growth curve models, and multilevel structural equation models, with each path labeled by its level of analysis. Results Momentary self-efficacy showed genuine within-person temporal dependence rather than independent scatter, and its trajectory was phase-linked to the competitive calendar and unevenly distributed across individuals. Between persons, fluctuation amplitude tracked the intensity of negatively valenced comparison rather than the sheer volume of feedback, and comparison orientation amplified that association on the negative arm alone. Within persons, negative comparison enlarged momentary instability both immediately and at a one-occasion lag, whereas positive signals steadied confidence only briefly. Amplitude was negatively associated with all three engagement facets beyond mean efficacy, most strongly for absorption, although a formal equality test separated absorption from dedication but not vigor. The direct reach of negative comparison into engagement proved facet-specific: vigor and absorption were affected, dedication was not, so the facet-by-facet hypothesis holds only in part. Sensitivity analyses adding competitive-calendar phase, splitting the series by season phase, partialling the person mean and the person-specific autoregression out of the amplitude index, entering competitive tier as a covariate, and checking participant-coded valence against automated sentiment scoring left the pattern intact. Discussion The findings extend social cognitive theory's vicarious and persuasive sources into a temporally dense digital regime and point toward protective coaching and self-regulation strategies.