Nonparametric regression analysis of longitudinal data (Q1188536)

In order to analyze longitudinal data the statistician is interested in the underlying curve, which produces the observed measurements. The classical method is parametric regression where a class of linear or nonlinear regression functions depending on a finite number of parameters is specified. Such a model is then fitted to the data by estimating the parameters, usually by the least squares method or the method of maximum likelihood. The main problem with parametric modelling is to search for a suitable model with not too many parameters, which gives a good fit to the data. Especially in biomedical applications this is often a very difficult task, so we have to look for an alternative to parametric regression. Such an alternative are nonparametric curve estimation procedures, e.g. kernel smoothing or kernel estimation. In contrast to parametric modelling, the assumptions on the function to be estimated are much weaker, namely only smoothness and differentiability requirements. In this monograph the author investigates several nonparametric curve estimation procedures, the emphasis being on kernel estimates as one of the most attractive methods of nonparametric regression, due to its simplicity, its computational advantages and its good statistical properties. He discusses the application of this method to longitudinal growth data and other longitudinal biomedical data. Questions of the choice of kernels and bandwidths or the estimation of derivatives are addressed. The basic approach is the estimation of each individual curve separately. Samples of curves can then be compared by means of ``longitudinal parameters''. In Chapter 2, the relation between longitudinal data, the fixed design regression model and other regression models is discussed. In Chapter 3, kernel estimates, weighted local least squares estimates, smoothing splines and orthogonal series estimates are reviewed. In Chapter 4, kernel and weighted local least squares estimators are studied more closely and the equivalence between these two methods is discussed. The kernel approach to the estimation of derivatives is described and some statistical properties are derived. Chapter 5 deals with the optimization of kernel and weighted local least squares methods w.r.t. various aspects of the choice of kernels. The order of the kernel determines the rate of convergence of the estimate and this is also reflected in finite sample studies. In Chapter 6, the kernel method is extended to the case of a multi- variate predictor variable, including the estimation of partial derivatives. Chapter 7 contains an overview over available methods for bandwidth choice. The choice of bandwidth is of special interest for derivatives and the question of local or global bandwidths. In Chapter 8, nonparametric estimates for peaks and zeroes and the joint asymptotic distribution of estimated location and size of peaks are discussed. Chapter 9 presents an application to the data of the Zürich longitudinal growth study of 45 boys and 45 girls. The superiority of nonparametric over parametric curve estimation can be demonstrated in this example. In Chapter 10, further techniques for the analysis of longitudinal medical data are summarized. In Chapter 11, sufficient conditions for local convergence in probability of linear curve estimators and for uniform convergence in probability and a.s. are compiled. In Chapter 12, FORTRAN computer programs are listed that carry out some of the procedures discussed in the text.