A0463
Title: Variable selection and estimation in non-linear mixed-effects models in high dimensional setting
Authors: Antoine Caillebotte - INRAE (France) [presenting]
Estelle Kuhn - INRA (France)
Sarah Sarah Lemler - Ecole CentraleSupelec (France)
Abstract: Mixed-effects models are a robust and increasingly popular tool for statistical modeling, in particular for the analysis of repeated measurements and longitudinal data. A non-linear mixed-effects model is considered, including high-dimensional covariates at the individual parameter modeling level. The focus is on variable selection and parameter estimation in this model. To handle the high dimensional setting and select a subset of relevant covariates, a LASSO type penalized maximum likelihood estimate is considered. The expectation maximization (EM) algorithm is a classical method for performing inference in mixed-effects models. However, it has several practical and theoretical limitations, such as the usual assumption that the model belongs to the exponential family. To circumvent this assumption, the use of the efficient Fisher-preconditioned stochastic gradient descent (Fisher-SGD) algorithm is proposed, which enables maximum likelihood inference in very general latent variable models. A proximal operator is added to this algorithm, which allows for penalized likelihood maximization and effective variable selection in high-dimensional contexts. The proposed algorithm's performance is illustrated in inference and variable selection through numerical simulations. The methodology is applied to a biological real dataset of wheat leaf senescence.
