A1059
Title: Estimation of threshold boundary regression models
Authors: ChihHao Chang - National University of Kaohsiung (Taiwan) [presenting]
Takeshi Emura - The Institute of Statistical Mathematics (Japan)
Shih-Feng Huang - National Central University (Taiwan)
Abstract: The threshold boundary regression (TBR) model is considered for sample splitting. The TBR model accommodates covariates in both the regression and threshold functions. The threshold function is allowed to be a nonlinear function of multiple covariates, constituting a hyperplane to describe data dynamics in two different states. TBR-WSVM, a two-stage method, is proposed that incorporates the weighted support vector machine (WSVM) and least-squares (LS) methods to estimate the TBR model. Under regularity conditions, the consistency of the TBR-WSVM estimators with their optimal convergence rates is evaluated. Several simulation experiments are conducted to investigate the finite sample performance of the TBR-WSVM estimator. Compared with two recently proposed methods, TBR-WSVM enjoys three advantages: (i) threshold parameters need not be prefixed with nonzero values, (ii) threshold parameter ranges need not be specified, and (iii) the threshold boundary can be non-linearly estimated. Finally, the TBR model is applied to real data analysis.
