Variable selection for random effects two-part models.

Area of Science:

• Statistics
• Biostatistics
• Econometrics

Background:

• Longitudinal studies often yield zero-inflated data (e.g., medical costs, alcohol consumption).
• High-dimensional data in these studies necessitates effective variable selection to avoid complexity.
• Existing variable selection methods are inadequate for sophisticated random effects two-part models.

Purpose of the Study:

• To develop a feasible variable selection method for random effects two-part models.
• To address the challenge of high dimensionality in zero-inflated longitudinal data.
• To adapt the minimum information criterion (MIC) for sparse estimation in these models.

Main Methods:

• Application of the minimum information criterion (MIC) for variable selection.
• Formulation of sparse estimation solvable using SAS Proc NLMIXED.
• Evaluation through simulation studies and a real-world alcohol dependence dataset.

Main Results:

• The proposed MIC-based method provides a practical approach to variable selection.
• Sparse estimation is effectively achieved and computationally feasible.
• The method demonstrates good performance in simulations and practical application.

Conclusions:

• The MIC method offers a viable solution for variable selection in random effects two-part models.
• This approach facilitates analysis of complex, high-dimensional, zero-inflated longitudinal data.
• The method is applicable to various fields, including health and social sciences.