Nonlinear Structural Functional Models

TL;DR: The aim of this dissertation is to ameliorate the shortcomings of existing methods through the development of unified nonlinear modeling procedures for the analysis of structural functional data through a general model-based framework.

Abstract: NONLINEAR STRUCTURAL FUNCTIONAL MODELS Michael R. Wierzbicki Wensheng Guo A common objective in functional data analyses is the registration of data curves and estimation of the locations of their salient structures, such as spikes or local extrema. Existing methods separate curve modeling and structure estimation into disjoint steps, optimize different criteria for estimation, or recast the problem into the testing framework. Moreover, curve registration is often implemented in a preprocessing step. The aim of this dissertation is to ameliorate the shortcomings of existing methods through the development of unified nonlinear modeling procedures for the analysis of structural functional data. A general model-based framework is proposed to unify registration and estimation of curves and their structures. In particular, this work focuses on three specific research problems. First, a Sparse Semiparametric Nonlinear Model (SSNM) is proposed to jointly register curves, perform model selection, and estimate the features of sparsely-structured functional data. The SSNM is fitted to chromatographic data from a study of the composition of Chinese rhubarb. Next, the SSNM is extended to the nonlinear mixed effects setting to enable the comparison of sparse structures across group-averaged curves. The model is utilized to compare compositions of medicinal herbs collected from two groups of production sites. Finally, a Piecewise Monotonic B-spline Model (PMBM) is proposed to estimate the locations of local extrema in a curve. The PMBM is applied to MRI data from a study of gray matter growth in the brain.