A SYSTEMATIC INVESTIGATION OF WITHIN-SUBJECT AND BETWEEN-SUBJECT COVARIANCE STRUCTURES IN GROWTH MIXTURE MODELS

The current study investigated how between-subject and within-subject variance-covariance structures affected the detection of a finite mixture of unobserved subpopulations and parameter recovery of growth mixture models in the context of linear mixed-effects models. A simulation study was conducted to evaluate the impact of variance-covariance structure difference, mean separation, mixture proportion and sample size on parameter estimates from growth mixture models. Data were generated based on 2-class growth mixture model framework and estimated by 1-, 2-, and 3-class growth mixture models using Mplus. Bias, precision and efficiency of parameter estimates were assessed as well as the model enumeration accuracy and classification quality. Results suggested that sample size and data overlap were key factors influencing the convergence rates and possibilities of local maxima in the estimation of GMM models. BIC outperformed ABIC and LMR in identifying the correct number of latent classes. Model enumeration using BIC could be improved by increasing sample size and/or decreasing overall data overlap, and the latter had more impact. Relative bias of parameters was smaller when subpopulation data were more separated. Both the magnitude of mean and variance-covariance separation and variance-covariance differences impacted parameter recovery. Across all conditions, parameter recovery was better for intercept and slope estimates than variance and covariances estimates. Entropy values were as high as the acceptable standards suggested by previous studies for any of the conditions even when data were very well-separated. Class membership assignment was more accurate when mean growth trajectories were more different among subpopulations and mixing proportions were more balanced