A1373
Title: Threshold-boundary Poisson regression models for counting data analysis
Authors: ChihHao Chang - National Chengchi University (Taiwan) [presenting]
Abstract: A threshold boundary Poisson regression (TBPR) model for analyzing count data with heterogeneous generating mechanisms is proposed. The proposed model allows the underlying regression structure to vary across subpopulations determined by an unknown threshold boundary, which can be either linear or nonlinear in the covariate space. To estimate the model parameters, an iterative two-stage estimation procedure is developed. In the first stage, a classification rule is constructed to partition the data into two groups based on their relative fitting performance under competing Poisson regression models. This classification rule is implemented via linear or nonlinear classifiers, enabling flexible modeling of complex threshold structures. In the second stage, separate Poisson regression models are fitted to each subgroup to update the regression parameters. Furthermore, the TBPR model is extended to accommodate zero-inflated count data, providing an alternative modeling framework to the conventional zero-inflated Poisson (ZIP) regression. Comprehensive simulation studies demonstrate that the proposed TBPR model achieves competitive or superior performance compared to the ZIP model, particularly when the data exhibit latent heterogeneity rather than excess zeros arising from a single mixture mechanism. An application to real-world data illustrates the practical effectiveness and interpretability of the proposed estimation algorithm.
