title 

Decomposing Variability in Functional Data Analysis

author

Theo Gasser and Daniel Gervini
University of Zuerich


abstract

This talk deals with three topics related to functional data analysis (FDA):

1. Concepts are explored how to assign and quantify variability in FDA
separately to amplitude variation and time variation. In the warping
approach, time variability is considered as nuisance; here, it is of similar
interest as classical amplitude variability. The method(s) are illustrated
with growth data and simulated data.

2. After registration of curves, FDA is usually based on smoothed PCA.
The resulting components are often
not well interpretable. Different schemes of decomposing amplitude
variability can lead to components with a better interpretation.

3. FDA via PCA leads to a multivariate vector of scores (per subject). A
more classical approach is feature extraction via a nonparametric or
semiparametric method. The two approaches will be compared for growth data
(Topic 3: if time permits).