EFFICIENT INFERENCE FOR LONGITUDINAL DATA VARYING-COEFFICIENT REGRESSION MODELS

Informative identification of the within-subject correlation is essential in longitudinal studies in order to forecast the trajectory of each subject and improve the validity of inferences. In this paper, we fit this correlation structure by employing a time adaptive autoregressive error process. Such a process can automatically accommodate irregular and possibly subject-specific observations. Based on the fitted correlation structure, we propose an efficient two-stage estimator of the unknown coefficient functions by using a local polynomial approximation. This procedure does not involve within-subject covariance matrices and hence circumvents the instability of calculating their inverses. The asymptotic normality of resulting estimators is established. Numerical experiments were conducted to check the finite sample performance of our method and an example of an application involving a set of medical data is also illustrated.