Are Bayesian regularization methods a must for multilevel dynamic latent variables models?

Area of Science:

• Psychometrics
• Statistical Modeling
• Longitudinal Data Analysis

Background:

• Intensive longitudinal data enables complex dynamic latent variable models.
• Challenges include overfitting, hierarchical structures, non-linearity, and sample size.
• Finite sample performance of priors (bias, accuracy, Type I error) requires attention.

Purpose of the Study:

• Compare Bayesian regularizing priors (ridge, Bayesian Lasso, adaptive spike-and-slab Lasso, regularized horseshoe).
• Evaluate their performance in multilevel dynamic latent variable models.
• Identify optimal priors for handling model complexity and estimation challenges.

Main Methods:

• Introduced a multilevel dynamic latent variable model.
• Conducted two simulation studies to assess prior performance.
• Performed a prior sensitivity analysis using empirical data.

Main Results:

• The ridge prior demonstrated effective sparse estimation without overshrinkage.
• Lasso and heavy-tailed priors underperformed compared to light-tailed priors in logistic models.
• Ridge priors offer a favorable balance between informativeness and generality.

Conclusions:

• Ridge priors are suggested for multilevel dynamic latent variable modeling.
• They effectively manage the trade-off between informativeness and generality.
• Avoid extreme shrinkage and heavy tails for improved model performance.