CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Penalized covariance smoothing and its impact on functional principal component analysis Authors:  Philip Reiss - University of Haifa (Israel) [presenting]
Abstract: Functional principal component analysis often proceeds by estimating the eigenfunctions of the covariance operator after smoothing an initial covariance function estimate. Penalized spline approaches to the covariance smoothing step have recently become popular. But the chosen penalization strategy can influence the final estimates to a degree that has generally been overlooked. In some cases, injudicious penalization can yield the false conclusion that a single eigendirection accounts for virtually all the variance. When the data represent developmental processes, such a conclusion may understate the diversity of developmental trajectories found in the population. This problem, and proposed solutions, are illustrated with data from a magnetic resonance imaging study of human cerebral cortex development.