A Comparison of Label Switching Algorithms in the Context of Growth Mixture Models

Area of Science:

• * Statistical modeling
• * Computational statistics
• * Psychometrics

Background:

• * Mixture models are frequently used in statistical analyses, but parameter estimation can be complicated by label switching.
• * Label switching, where latent classes are incorrectly assigned across replications, is a significant challenge in mixture model simulations.
• * Various methods exist to correct for label switching, but their comparative performance is not fully understood.

Purpose of the Study:

• * To compare the effectiveness of different label switching correction methods in mixture models.
• * To evaluate both a priori (identifiability constraints, algorithm training) and post hoc algorithms.
• * To assess how factors like the number of latent classes, class probabilities, and class separation influence correction accuracy.

Main Methods:

• * A simulation study was conducted using a growth mixture model.
• * The simulation design included 18 conditions by crossing three variables: number of latent classes, latent class probabilities, and class separation.
• * The accuracy of a priori identifiability constraints, a priori algorithm training, and four post hoc algorithms (Tueller et al., Cho, Stephens, Rodriguez & Walker) was tested.

Main Results:

• * A priori algorithm training demonstrated the highest classification accuracy across all tested conditions.
• * The Rodriguez and Walker post hoc algorithm was effective for aggregating class output when class order was not critical.
• * All tested post hoc algorithms improved classification accuracy over baseline, with optimal performance in two-class models with high class separation.
• * Combining a priori class constraint algorithms with post hoc methods enhanced classification accuracy.

Conclusions:

• * For mixture model simulations, a priori algorithm training is the most reliable method for accurate classification.
• * Post hoc algorithms offer valuable improvements, particularly in simpler models with distinct classes.
• * A combination of a priori and post hoc methods is recommended for robust label switching correction.