A0304
Title: A boosting method for variance components selection in linear mixed models
Authors: Michela Battauz - University of Udine (Italy) [presenting]
Paolo Vidoni - University of Udine (Italy)
Abstract: Boosting is a method developed in machine learning and later translated to statistical analysis to estimate the parameters of a model. The procedure iteratively improves the fit of the model by updating a subset of parameters at each step. In this approach, early stopping is fundamental to perform model selection and prevent overfitting. The proposals in the literature for random effects models focus on the fixed part of the model, while the variables with random effects should be pre-specified. We present a novel method to select the variance components of the model that considers the negative profile log-likelihood as the objective function to minimize. The issue of non-convexity of such a function is overcome by exploiting the directions of negative curvature, which allows scaping saddle points or local maxima. Simulation studies show the good performance of the proposal in detecting the real structure of the data, while an application to the analysis of total nitrate concentration further illustrates the procedure.
