Document Type : Original Paper
Authors
Department of Statistics, Tarbiat Modares University, Tehran, Iran.
Abstract
The study addresses the challenges of analyzing time-to-event data, particularly emphasizing the discrete nature of durations, such as the number of years until divorce. This frequently results in zero-inflated survival data characterized by a notable frequency of zero observations. To address this, the study employs the zero-inflated discrete Weibull regression (ZIDWR) model, which serves as a suitable framework for evaluating the impact of explanatory variables in survival analysis. However, challenges such as nonstationarity in the relationship between variables and responses and spatial heterogeneity across geographical regions can result in a model with too many parameters To mitigate this, we propose a spatial clustering approach to summarize the parameter space. This Paper leverages nonparametric Bayesian methods to explore the spatial heterogeneity of regression coefficients, focusing on the geographically weighted Chinese restaurant process (gwCRP) for clustering the parameters of the ZIDWR model. Through simulation studies, the gwCRP method outperforms unsupervised clustering algorithms clustering K-means and the standard Chinese restaurant process (CRP), exhibiting superior accuracy and computational efficiency, particularly in scenarios with imbalanced cluster sizes. This improved performance is quantitatively demonstrated through higher Rand indices, lower average mean squared error (AMSE) in parameter estimation and superior log pseudo-marginal likelihood (LPML) values. Applying this methodology to Iranian divorce data reveals distinct spatial clusters characterized by varying covariate effects on the probability of divorce within the first five years of marriage and the subsequent time to divorce.
Keywords
Main Subjects
)