Statistical methods for large n and p problems

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Abstract Modern observational and experimental biological data has undergone a revolution. Driven by new biotechnology and computing advances, high dimensional, high density, functional multilevel and longitudinal biological signals are becoming commonplace in medical and public health research. These types of signals historically occurred in small clinical or experimental settings, often referred to as the "small n, large p" problem. We view the extension of these biological signals to cohort studies with longitudinal or hierarchical structure as a next generation of biostatistical problems. We've taken to calling this the "hierarchical large n, large p" problem. The goal of this grant is to introduce general methods for analyzing this form of biostatistical data. We propose three major aims for the analysis of multilevel or longitudinally collected biosignals. The first extends multilevel functional principal components, the investigators' generalization of functional principal components, to longitudinal and high dimensional settings. The second considers the investigators bi-directional filtering and extends it in high-dimensional and longitudinal settings. The third considers model-based independent component blind source separation and extends it to longitudinal settings. To solve this aim, we will also consider the fundamental problem of running MCMC samplers for high dimensional parameter spaces. Specifically, no current work exists for convergence control when the number of parameters is larger than the number of iterations. We propose a method of convergence control using finite population sampling. Our methods will be applied to unique data sets involving imaging (MRI, fMRI, DTI), electrophysiology (EEG, ECOG), sleep measurement (polysomnography) and novel measurements of aging (accelerometer). In the preliminary results, we demonstrate our capacity for working with such data with novel findings in the analysis of EEG, MRI and fMRI data sets. Methods such as unsupervised clustering, blind source separation and dimension reduction are generally recognized first steps in analyzing high dimensional data, and have been applied success- fully in an amazingly diverse collection of settings. Our proposal generalizes these basic approaches to high dimensional data while considering hierarchical and longitudinal directions of variation. Hence, our approaches will form a basic foundation for next generation biomedical functional data. PUBLIC HEALTH RELEVANCE: Modern observational data is often longitudinal or multilevel functional biological signals. We propose a foundational approach for the analysis of such data, including scalable computing to next generation data sets.