An Alternative Approach to Growth Mixture Model: Performance of Covariance Pattern Mixture Model

This study evaluates the performance of five modeling approaches—unconstrained growth mixture model (GMMU), constrained GMM (GMMC), latent class growth model (LCGM), covariance pattern mixture model (CPMM) with compound symmetry (CPMM-CS), and CPMM with Toeplitz (CPMM-TP) structure—in identifying heterogeneous growth trajectories in longitudinal data. A Monte Carlo simulation was conducted using a three-class growth mixture model with five time points as the population model, varying sample size, class separation, and class proportion disparities. Model performance was assessed based on convergence rate, class enumeration accuracy, and parameter estimation accuracy. The results indicate that GMMU exhibited the lowest convergence rates, particularly under small sample sizes and low class separation. CPMM-TP achieved the highest class enumeration accuracy and outperformed GMMU in both convergence and classification, especially with small samples where GMM typically struggled. While CPMM-CS showed stable convergence and accurate parameter estimation, its class enumeration performance was comparatively lower. LCGM and GMMC demonstrated significant bias in slope and intercept estimates when class separation was low. These findings suggest that CPMM-TP provides a viable alternative to GMM, offering improved convergence stability and class enumeration accuracy, which are critical for the reliable identification of underlying trajectories.