A1724
Title: Graph-based change-point detection for regression changes under covariate shift
Authors: Seho Kim - Iowa State University (United States)
Lynna Chu - Iowa State University (United States) [presenting]
Abstract: Change-point detection in regression aims to determine whether and when the conditional relationship between covariates and response changes over time. Classical structural break tests typically rely on strong global parametric assumptions that permit extrapolation across covariate regimes, but they can be sensitive to model misspecification. Modern nonparametric methods are more flexible, yet many target changes in the joint distribution of the data and may therefore fail to distinguish conditional distributional change from covariate shift. A graph-based, covariate-aware framework for detecting conditional changes in regression is proposed that accommodates time-varying covariate distributions through local comparability. The resulting test is designed to isolate changes in the conditional mechanism from changes in the covariate distribution itself. Its asymptotic null distribution is derived and simulations demonstrate that the method performs well across a wide range of settings.
