On prediction of future observation in growth curve model.

Abstract

Rao proposed and compared several approaches for predicting future observations in a growth curve model. The assessment of associated prediction efficiency for different prediction methods were evaluated by Cross-Validation Assessment Error (CVAE). He used three data sets, each with a limited number of subjects (13-27) and also with a limited number of repeated measurements (4-7) per subject, to illustrate the prediction methods. In the present paper, we applied four of the prediction methods discussed by Rao, on a data set with a relatively large number of subjects (174) and also with a larger number of measurements (21) per subject, using the polynomial function and log-linear function. We propose to use the restricted cubic spline function as an alternative growth curve model and compare its performance with the polynomial function and log-linear function. It turns out that, at least for larger data sets such as that used in this paper, the prediction methods perform somewhat better when the growth is described by restricted cubic spline function than when the growth is described by polynomial function and log-linear function.