Conditional association structure among general self-efficacy, self-esteem, and emotion reactivity in university students: evidence from Gaussian graphical and Bayesian network analyses

General self-efficacy (GSE), self-esteem (SE), and emotion reactivity are associated with university students’ psychological adjustment. Conventional regression models are useful for testing prespecified associations, whereas network analysis offers a complementary exploratory representation of conditional associations among multiple constructs. This study examined the conditional association structure among GSE, SE, and three emotion-reactivity dimensions using Gaussian graphical and Bayesian network models. Data were obtained from 1,256 Chinese university students recruited through class-level convenience sampling. The primary network analyses used 1,253 complete cases. Because the Rosenberg Self-Esteem Scale (SES) showed a pronounced wording structure, we conducted a dedicated measurement analysis comparing a conventional one-factor model, a correlated positive/negative wording-factor model, and a general-factor-plus-negative-wording-method-factor model. We then repeated both GGM and BN analyses under positive- and negative-wording SES scoring to evaluate measurement dependence. All analytically derived quantities after data entry are generated from the deposited item-level file; the original counts of 1,345 returned questionnaires and 87 initially excluded questionnaires are historical study-flow records and cannot be regenerated from the deposit. The strongest edges were among the three emotion-reactivity subscales. The ten-item SES did not support an unproblematic one-dimensional self-esteem interpretation: a one-factor measurement model fitted poorly, and models allowing wording structure fitted substantially better. Accordingly, the full-score GSE-SE estimate (regularised partial r = 0.31) is reported only with an explicit measurement-dependence qualification. GSE-ER2 ( r = 0.07) was the strongest positive cross-domain edge not involving the problematic SES composite. BN sensitivity analyses showed that SE-related skeleton and CPDAG features changed across wording-block representations; in particular, the compelled SE-to-GSE orientation was specific to representations in which the GSE-SE edge and the corresponding v-structure were retained. Self-cognition and emotion reactivity were conditionally associated in this sample, although cross-community edges were weak relative to the strong within-scale ERS subscale edges. The observed pattern of conditional associations, including the weak GSE-ER2 and SE-ER1 cross-community connections, warrants testing in longitudinal and experimental designs; the present cross-sectional network does not identify causal mechanisms or established intervention targets. Internal consistency was low for the self-esteem measure in this sample, and the self-esteem findings should be read with that measurement limitation in mind. Not applicable.