Abstract: Revisiting the Model Size Effect in Structural Equation Modeling (SEM)
Dexin Shi,Taehun Lee,R. Terry
DOI: https://doi.org/10.1080/00273171.2014.989012
2015-01-02
Multivariate Behavioral Research
Abstract:Revisiting the Model Size Effect in Structural Equation Modeling (SEM) Dexin Shi, Taehun Lee, and Robert A. Terry Department of Psychology, University of Oklahoma It is generally believed that fitting a large SEM model with moderate to small sample sizes results in a likelihood ratio statistic distribution with inflated Type I error rate, thus poorly approximating the asymptotic chi-square distribution. This bias in terms of empirical likelihood ratio statistic distribution is known as the model size effect, and is considered an important issue in psychological studies. The model size issue has been investigated in previous studies (Herzog, Boomsma, & Reinecke, 2007; Moshagen, 2012). However, the notion of “model size” among different studies is ambiguously defined. In previous literatures, size of the model was indicated by different indexes, including the number of observed variables (p), the number of parameters to be estimated (q), degrees of freedom (df = p (p +1)/2 – q), and ratio of the observed variables to latent factors (p/f ). In the current study, we explored the essential influential factors for the model size effects using Monte Carlo simulation. Our design differed from previous studies in the following two ways. First, the factors of p/f , p, and q were manipulated independently; when changing one factor, we kept other factors constant. Second, we manipulated the values of q to a greater degree by specifying more parameters to be estimated (e.g., cross loadings or correlated residuals) and (or) constraining the parameters to their true population values (to decrease q). The results are summarized as follows. First, bias in the likelihood ratio statistic is not substantially influenced by the number of latent factors (f ), or the ratio of the observed variCorrespondence concerning this abstract should be addressed to Dexin Shi, Department of Psychology, University of Oklahoma, Norman, OK 73019-2007. E-mail: dexin.shi@ou.edu ables to latent factors (p/f ). Second, model size effects are associated with the number of observed variables (p). Generally, larger p is associated with larger bias. Third, model size effects also appear to be a function of the number of estimated parameters (q), such that increasing the number of estimated parameters results in smaller bias. Finally, given that both p and q are influential factors of the model size effects, and their effects act in different directions, we further tested whether the effects of p and q can be explained more simply by the degrees of freedom (df ). Results indicate that p and q still have unique effects, even holding df constant. The present study expands upon conclusions from previous studies. Although p was found to be most important determinant of model size effects, we also demonstrated that the effect of q is not ignorable after controlling the influence of other factors. In addition, the combined effects of p and q cannot be fully explained as a simple function of df . Our findings also provide support for empirical researchers to resolve the issue of model size effects by implying relevant corrections, when developed, must consider both p and q in their formulation.