Information and other criteria in structural equation model selection

Publisher Website
Google Scholar
Abstract
This article presents the results of a simulation study evaluating information criteria in conjunction with other well-known criteria for model selection in structural equation modeling (SEM). Two sets of simulation experiments were performed. In both sets, sample sizes of n = 100,400,1000,6000 were used and the performance of 18 criteria was assessed by the frequency with which each of five analytic models was selected as best by each criterion in 500 replications. In the first set of experiments correctly specified analytic models (noncentrality parameter 0) were entertained in combination with misspecified ones, while in the second set all five models were misspecified. In both sets of experiments, we found that the information criteria perform better than the other criteria overall, but that Cudeck and Browne's cross-validation index ( CVI) remains an attractive option. Within the class of information criteria, Akaike's information criterion (AIC) is found to show some overfitting tendency. We demonstrate -analytically and through our simulations - the difference between the asymptotic behavior of the AIC and other information criteria. Contrary to suggestions in the literature, AICs overfitting tendency is not found to be markedly stronger for the larger sample sizes.