A0174
Title: Post-selection inference for fixed and mixed parameters in linear mixed models
Authors: Katarzyna Reluga - University of Bristol (United Kingdom)
Gerda Claeskens - KU Leuven (Belgium)
Stefan Sperlich - University of Geneva (Switzerland) [presenting]
Abstract: While post-selection inference has received considerable attention in linear models, it is a neglected topic in the field of mixed models and mixed effect prediction. We developed methods and asymptotic theory for post-selection inference when the conditional Akaike information criterion was employed for model selection in a linear mixed model. These are used to construct confidence intervals for regression parameters, linear statistics and mixed effects under different scenarios, namely nested and general model sets as well as sets of misspecified models. The theoretical analysis is accompanied by simulation studies that confirm good performances. Moreover, they reveal a startling robustness of the classical confidence intervals for mixed parameters, which is in strong contrast to the findings for fixed parameters, indicating that random effects could automatically adjust for model selection. We illustrate our methodology along a study of the body mass index across different clusters in the US
