Variable selection in a linear growth curve model with autoregressive within-individual errors

Abstract

In this paper, we consider variable selection schemes in a linear random coefficient growth curve model. Unbalanced within-individual data is assumed to have first-order linear correlation (p). A variable selection scheme suggested in the between-individual model is a weighted stepwise selection. A sim-ulation study is conducted to compare the performances of the maximum likelihood (ML) and the ordinary least square (OLS) estimators in the within-individual regression model in terms of the ability of variable selection in the between-individual model. Simulation results indicate the following: Under the suspicion of high autocorrelation error (p>0.5), the ML estimation is necessary in the within-individual model. When it is believed that the p is as small as 0.1 and the heterogeneity of variance is relatively low then, the OLS estimator can replace the ML estimator in terms of selection performance in the between-individual model.