Title: Penalty method for variance component selection
Authors: Hua Zhou - UCLA (United States)
Juhyun Kim - University of California, Los Angeles (United States) [presenting]
Abstract: Variance components models, also known as mixed effects model, are central themes in statistics. When there is a large number of variance components, one may wish to select a subset of those that are associated with response. Existing methods are limited to finding random components at individual level or within one variance component. We propose selection of variance components based on a penalized log-likelihood with adaptive penalty. This is achieved via a majorization-minimization (MM) algorithm, which is well known for being simple, numerically stable, and easy to implement. Performance of the proposed methodology is shown empirically through simulation studies and real data analysis. In theory, we establish a non-asymptotic error bound for the output from the algorithm and characterize the region in which the MM iterates converge to a global optimum of the population likelihood. This result provides a theoretical guideline in terminating MM iterations.