A1370
Title: Efficient and distribution-free variable selection in spatial models with zero inflation
Authors: Chun-Shu Chen - National Central University (Taiwan) [presenting]
Chia-Ming Hsu - National Central University (Taiwan)
Abstract: Extreme climate events, such as heavy rainfall, are often characterized by strong spatial dependence and a substantial proportion of zeros due to their rarity and localized occurrence. Modeling such data poses significant challenges, as it requires simultaneously accommodating spatial dependence, excess zeros, and complex covariate effects. A spatial modeling framework with zero inflation is utilized, with a particular emphasis on developing an efficient and distribution-free variable selection procedure for extreme climate analysis. The core contribution is a new distribution-free selection criterion that avoids reliance on a fully specified likelihood, thereby enhancing robustness against model misspecification, while remaining computationally efficient and scalable in high-dimensional spatial settings. Inspired by the regularization principle of Lasso, the proposed method enables fast and stable identification of influential covariates while preserving modeling flexibility. Extensive simulation studies demonstrate its accuracy, robustness, and computational efficiency. An application to Taiwan's 2016 daily extreme rainfall data further illustrates its practical value for environmental risk assessment and climate-related decision-making.
