Nonlinear Spatial Prediction with Predictor Subject to Limit of Detection

• Activity Number: 131 - Methods for Spatial, Temporal, and Spatio-Temporal Data
• Type: Contributed
• Date/Time: Monday, August 9, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Minho Kim* and Kyuhee Shin and Gyuwon Lee and Joon Jin Song
• Companies: Baylor University and Kyungpook National University and Kyungpook National University and Baylor University
• Keywords: Multiple Imputation; Limit of Detection; Censoring; Generalized Additive Model; Regression Kriging; Spatial Prediction

Abstract

Accurate spatial predictions are the primary interest in many studies, but variables subject to limit of detection (LOD) make it difficult to achieve the goal. In spatial prediction, considerable efforts have been made in the literature for censored observations, however, much less has been explored when a predictor is subject to censoring. Moreover, most spatial prediction studies assume a linear relationship between observations and predictors, which may be too restrictive in some applications. To fill these gaps, we propose a novel hybrid model, which utilizes generalized additive model (GAM) for modeling the nonlinear relationship and ordinary kriging, for spatial prediction. We incorporate a multiple imputation (MI) based on conditional expectation into the proposed model to handle the censored observations. A simulation study is performed to assess the predictive performance and robustness of the model and compared to other frequently used imputation methods in practice. The proposed model is also illustrated in quantitative precipitation estimation (QPE) with rain gauge and radar in South Korea.