Shrinkage estimation in linear mixed models for longitudinal data (Q723454)

A linear mixed model is considered where some of the fixed effect parameters are under a linear restriction. The random effects are treated as nuisance parameters. Six estimators are studied: unrestricted and restricted maximum likelihood estimators (UE and RE), pretest estimator (PE), non-penalty positive shrinkage estimator (PSE), and two penalty estimators (Lasso and adaptive Lasso). A simulation study shows the following: PSE performs better than the penalty estimators when there are many inactive covariates in the model; the penalty estimators perform better than PSE when the number of inactive covariates is small; RE performs best at or near the restriction, but it is dominated by PSE as one moves away from the restriction; PE, PSE, and penalty estimators all outperform UE.