| Popis: |
Class enumeration and detection of latent heterogeneity when fitting the Growth Mixture Model (GMM) are multifaceted processes that are influenced by several factors, resulting in invalid class-structure, mainly estimated by Maximum Likelihood using Expectation-Maximization algorithm (ML/EM). A recent automated data mining algorithm, Structural Equation Modelling Forest (SEM Forest), is a promising alternative. The current study investigated the performance of the two estimators (i.e., ML/EM and SEM Forest) in enumerating classes correctly when fitting one-class and three-class GMM under various experimental conditions by conducting two Monte Carlo studies. Several design factors were manipulated (i.e., class separation, sample size, mixing proportions, inclusion of covariates, and within-class normality). The findings of two studies revealed a remarkable divergence in the two estimators’ performances under the normal and nonnormal cells. The findings showed that the two estimators had similar high enumeration rates, IC’s enumeration accuracy, acceptable relative bias, and low MSE under the normal conditions. This good performance scaled up as the sample size and separation increased. However, the two estimators failed to keep this good performance under the nonnormal cells, as demonstrated by lower enumeration rates, lower IC’s accuracy, more prominent relative bias, and higher MSE. Several unanticipated trends related to the influences of design factors on the two estimators’ performances were noted under nonnormal cells, precisely trends pertain to the sample size and class separation. This study had fruitful implications and provided applied researchers with detailed descriptions about the performance of the two methods, building a bridge between confirmatory analyses and exploratory algorithms. |
