Evaluation of Generative Adversarial Imputation Nets’ Performance in Handling Missing Data in Structural Equation Modeling
Luqi He, Yingke Lu, Carl F. Falk, Heungsun Hwang · Structural Equation Modeling A Multidisciplinary Journal · 2025
Missing data are a common challenge in structural equation modeling (SEM), potentially leading to biased estimates and reduced power. Full information maximum likelihood (FIML) and multiple imputation (MI) are widely used to address this issue. Recently, generative adversarial imputation nets (GAIN), a machine learning–based method, have shown promise under high missingness; however, their utility within SEM contexts remains largely unexplored. This simulation study compared GAIN, FIML, and MI across several experimental factors. Under correct model specification, all methods yielded comparable estimates, with GAIN exhibiting greater variability and poorer recovery of model fit. Under model misspecification, performance differences became more pronounced with increasing missingness. Below 50%, all methods performed similarly, though GAIN showed higher variability and poorer fit recovery. At 50%, MI and GAIN outperformed FIML with comparable accuracy. At 75%, GAIN produced more accurate estimates than MI but continued to show the greatest variability and poorest model fit recovery.