Bayesian nonparametric centered random effects models with variable selection.

Area of Science:

• Statistics
• Biostatistics
• Computational Biology

Background:

• Parametric assumptions (e.g., normal distribution) for random effects in linear mixed effects models can lead to errors in variable selection.
• Nonparametric models with nonzero mean random effects face identifiability issues for fixed effects.
• Existing methods may struggle with accurate variable selection when normality assumptions are violated.

Purpose of the Study:

• To develop a Bayesian variable selection method for linear mixed effects models.
• To address challenges posed by non-normal and potentially nonzero mean random effects.
• To improve the accuracy and interpretability of model selection in complex data.

Main Methods:

• Utilized a Bayesian approach with Dirichlet process priors for nonparametric characterization of subject-specific random effects.
• Developed a method to simultaneously resolve bias and model the conditional distribution of random effects flexibly.
• Employed a stochastic search Gibbs sampler for efficient identification of relevant fixed and random effects.

Main Results:

• The proposed Bayesian method effectively handles violations of normality assumptions in random effects.
• The approach successfully addresses identifiability problems associated with nonzero mean random effects.
• Simulations demonstrated superior performance compared to existing variable selection techniques.

Conclusions:

• The Bayesian nonparametric approach offers a robust solution for variable selection in linear mixed effects models.
• This method enhances model interpretability and accuracy, particularly when standard assumptions are not met.
• The approach is validated through simulations and a real-world application in a rodent bioassay.