Varying-smoother models for functional responses

This paper studies estimation of a smooth function $f(t,s)$ when we are given functional responses of the form $f(t,\cdot)$ + error, but scientific interest centers on the collection of functions $f(\cdot,s)$ for different $s$. The motivation comes from studies of human brain development, in which $t$ denotes age whereas $s$ refers to brain locations. Analogously to varying-coefficient models, in which the mean response is linear in $t$, the "varying-smoother" models that we consider exhibit nonlinear dependence on $t$ that varies smoothly with $s$. We discuss three approaches to estimating varying-smoother models: (a) methods that employ a tensor product penalty; (b) an approach based on smoothed functional principal component scores; and (c) two-step methods consisting of an initial smooth with respect to $t$ at each $s$, followed by a postprocessing step. For the first approach, we derive an exact expression for a penalty proposed by Wood, and an adaptive penalty that allows smoothness to vary more flexibly with $s$. We also develop "pointwise degrees of freedom," a new tool for studying the complexity of estimates of $f(\cdot,s)$ at each $s$. The three approaches to varying-smoother models are compared in simulations and with a diffusion tensor imaging data set.