Improving convergence in growth mixture models without covariance structure constraints.

Area of Science:

• Statistics
• Biostatistics
• Psychometrics

Background:

• Growth mixture models (GMMs) are widely used to identify unobserved heterogeneity in developmental trajectories.
• Nonconvergence is a common issue when fitting GMMs to empirical data, often leading to model misspecification.
• Modifying random effect covariance structures to achieve convergence can negatively impact parameter estimates and class assignments.

Purpose of the Study:

• To advocate for marginal modeling approaches in GMMs to avoid convergence problems.
• To demonstrate the importance of retaining class-specific covariance structures.
• To present covariance pattern growth mixture models (CP-GMMs) as a viable alternative.

Main Methods:

• A simulation study was conducted to compare the performance of different GMM specifications.
• The study contrasted models with constrained versus unconstrained random effect covariance structures.
• Covariance pattern growth mixture models were explored as a marginal modeling alternative.

Main Results:

• Constraining random effect covariance structures improves model convergence but compromises parameter estimates, class assignments, and class enumeration.
• Marginal models, specifically CP-GMMs, facilitate convergence without sacrificing essential class-specific covariance information.
• CP-GMMs offer a robust approach, especially with smaller sample sizes and data attrition.

Conclusions:

• Retaining class-specific covariance structures is crucial for accurate GMM results.
• Marginal modeling approaches, such as CP-GMMs, provide a practical solution to convergence issues in GMMs.
• CP-GMMs offer a promising method for analyzing complex growth trajectories in real-world applications, including PTSD research.