Studies on semi-parametric varying coefficient models

Abstract

In many economic and geographic studies, we may have spatially referenced covariates providing information about the spatial distribution that impacts the response variable. The spatial varying coefficient model (SVCM) has been an effective tool for exploring such information by modeling spatial nonstationarity. In this thesis, we study the SVCM and address several challenges in estimating the varying coefficient functions over complex domains in different scenarios. In chapter 2, we consider a new class of semi-parametric regression models called the generalized partially linear spatially varying coefficient model (GPLSVCM). We propose using the bivariate penalized spline over triangulation (BPST) method to approximate the coefficient functions and employing a quasi-likelihood maximization to obtain model estimators. The proposed method can handle data distributed over arbitrarily shaped domains with complex boundaries and interior holes. We prove the consistency of the estimators under some regularity conditions. Additionally, we propose a model selection procedure via BIC that can accurately identify the covariates with constant and varying effects. In chapter 3, we introduce a new R package GPLSVCM, which integrates model structure identification, variable selection, model fitting, and predictive inference for GPLSVCMs. To account for high-dimensional data, we propose a doubly penalized approach for simultaneous variable selection and model structure identification. The proposed method can efficiently remove irrelevant covariates while detecting constant and varying components of the coefficients. To quantify the uncertainty in a single prediction, we propose three resampling-based methods for constructing prediction intervals that attain target coverage probability. Compared with existing R packages for SVCMs, GPLSVCM is more flexible and computationally cheaper, so it can be widely applied in spatial data analysis over any arbitrarily shaped domain. In chapter 4, we develop a new volatility model by allowing spatially varying coefficients in spatial GARCH models. This model captures volatility behaviors over space and investigates the relationship between some explanatory variables and the volatility at each location. A two-stage quasi-likelihood maximization via BPST is developed to estimate the model over a complicated domain. For each chapter, we conduct both simulation studies and real-data applications to demonstrate the performance of our approach.