An investigation of missing data handling methods in terms of model fit indices in path analysis

This research aims to compare the effects of various missing data handling methods—Listwise Deletion (LWD), Regression Imputation (RI), Series Mean (SM), Multiple Imputation (MI), and Expectation Maximization (EM)—on model fit indices within the framework of path analysis. Under the assumption of a Missing Completely at Random mechanism, datasets with 5%, 10%, and 20% missing data were analyzed across five proficiency levels (1a, 1b, 2, 3, and 4) using sample sizes of 200, 500, and 1000 participants. Model fit indices, including χ², GFI, RMSEA, CFI, SRMR, NNFI, and AGFI, were calculated for each completed dataset and compared to the values obtained from the original (complete) dataset. Based on PISA 2022 data, the relationships between creative thinking and achievement in mathematics, reading, and science were examined through path modeling. The findings indicated that EM consistently produced the most favorable model fit values across all sample sizes and levels of missingness. RI generated results that were comparable to EM, particularly in larger samples. While SM demonstrated relatively stable performance at low levels of missing data, its effectiveness declined as the missingness increased. MI failed to provide sufficient model fit in certain conditions. The poorest results were observed with LWD due to the loss of substantial amounts of data. Overall, as the proportion of missing data increased, model fit indices tended to deteriorate.