Sparse Variational Analysis of Linear Mixed Models for Large Data Sets.

Area of Science:

• Statistics
• Computational Biology
• Epidemiology

Background:

• Longitudinal and multi-level data with many predictors or large sample sizes are increasingly common.
• Existing mixed-effects models struggle with computational demands and inference in these settings.
• Subset selection and sparse data characterization are desirable for high-dimensional data.

Purpose of the Study:

• To develop a computationally efficient Bayesian method for mixed-effects models with large numbers of predictors or large sample sizes.
• To enable sparse selection of random effects in high-dimensional data.
• To provide a fast approximation to the posterior distribution.

Main Methods:

• A novel variational Bayes approximation to the posterior distribution is proposed.
• The method approximates the posterior of variance component parameters.
• Priors are chosen to induce sparsity and facilitate random effect selection.

Main Results:

• The proposed variational method offers a computationally feasible alternative to Markov chain Monte Carlo (MCMC) for large-scale problems.
• The method demonstrated effectiveness in simulation studies.
• The approach was successfully applied to an epidemiological dataset.

Conclusions:

• The variational Bayes approach provides a fast and effective solution for fitting mixed-effects models to large, high-dimensional datasets.
• This method facilitates sparse random effect selection, improving model interpretability.
• The technique is suitable for complex data structures encountered in various scientific fields.