Parameter cascades and profiling in functional data analysis (Q964612)

A penalized least squares B-spline curve fitting procedure is considered with the penalty term of the form \(\lambda\int(L x(t))^2dt\), where \(L=L_\beta\) is some linear differential operator, and \(\lambda\) is the smoothing parameter. The form of \(L_\beta\) is taken to be dependent on the general trend deduced from the data. E.g., if a superposition of linear and sinusoidal trends \(x_0(t)=w_1+w_2t+w_3\sin(e^\beta t)+w_4\cos(e^\beta t)\) is imposed then \(L_\beta x=e^{2\beta}x''+x''''\) satisfies \(L_\beta x_0=0\). This technique deals with three types of parameters: local parameters \(c\) which are the spline coefficients, structural parameters \(\beta\), and the complexity parameter \(\lambda\). All the parameters are estimated via minimization of some criteria. The idea of the article is that these criteria should be different: the penalized least squares (LS) for \(c\) (given \(\beta\) and \(\lambda\)), the unpenalized LS for \(\beta\) with \(\hat c=\hat c(\beta)\) being previously estimated, and the cross-validation criterion is applied to choose \(\lambda\). This technique is called a parameter cascade. The variances of the parameter estimates are approximated via the delta method. Results of simulations and an application to melanoma incidence data in Connecticut are presented.