← Back to Vol. 11 Issue No. 01 — March
Abstract
This Monte Carlo study reexamines the performance of 13 commonly used fit indices in structural equation modeling (SEM) across a range of model conditions. Simulated data were generated from both confirmatory factor analysis (CFA) and latent curve autoregressive moving average (LCARMA) models under correct and misspecified conditions, with sample sizes of 125, 250, 500, 1000, 2000, and 4000, and varying degrees of model misfit. The results reveal that widely cited cutoff values—such as .95 for CFI and .08 for SRMR—often failed to distinguish correct from incorrect models, particularly in small samples or under mild misfit conditions. RMSEA demonstrated the most consistent performance, generally supporting a .05 threshold across scenarios. In contrast, incremental indices such as CFI and Gamma required values approaching .99 to achieve acceptable specificity. Parsimony indices (PGFI and PNFI) exhibited a systematic bias favoring complex models, even when model fit was held constant. The severity of misspecification and the complexity of the model structure both substantially influenced the effectiveness of fit indices. These findings challenge the application of universal cutoff values and underscore the need for context-sensitive interpretation of fit in SEM practice.