Psychometric structure and score interpretation of the SCL-90 among Chinese first-year undergraduates: an analysis of an open dataset

Introduction The Symptom Checklist-90 (SCL-90) provides total and domain scores whose shared and distinct information remains debated. Methods We analyzed an open dataset of 14,371 Chinese first-year undergraduates in calibration and validation halves. Four item-level structures (unidimensional, nine correlated factors, second-order, and orthogonal bifactor) were compared on the same 83 domain-scored items using robust maximum likelihood (MLR) in both halves and mean- and variance-adjusted weighted least squares (WLSMV) for ordered items in validation. A separate single-factor model included all 90 items. Evaluation combined absolute and relative fit, numerical diagnostics, loadings, residuals, and general and specific reliability. Results For the 83-item comparison, the correlated-factor model had a non-positive-definite latent matrix. The bifactor model improved on the admissible second-order model, but its WLSMV comparative fit index (CFI) advantage was only 0.0079; MLR validation fit remained mixed (CFI = 0.849, Tucker-Lewis index [TLI] = 0.842, root mean square error of approximation [RMSEA] = 0.045, standardized root mean square residual [SRMR] = 0.042). Under that MLR bifactor specification, general explained common variance (ECV) was 0.813 [0.804, 0.822] and omegaH was 0.963 [0.961, 0.965]. Specific omegaHS ranged from 0.017 to 0.293, with somatization, phobic anxiety, and hostility retaining comparatively more specific information. WLSMV general omegaH was 0.967 and specific omegaHS ranged from 0.005 to 0.186. The 90-item single-factor model showed limited incremental fit (MLR CFI = 0.769/0.773; WLSMV CFI = 0.896); RMSEA was 0.052/0.052 and 0.044, and residual dependence persisted. High 90-item alpha (0.978) therefore coexisted with departures from strict unidimensionality. An exploratory mixture audit yielded no stable class solution. Discussion These data support a substantial general-severity component and uneven domain-specific information within the modeled 83-item structure. Its reliability decomposition does not establish that of the conventional 90-item total or global severity index. Independent-sample replication remains necessary.