Modelos espaciais com distribuição normal independente

Abstract

Linear spacial models, known as geostatistics models, are used in the study of phenomena that are spatially correlated. Although it is usual to admit that the spacial process follows a Gaussian distribution, the presence of atypical observations can make this analysis inappropriate. Thus, it might be necessary to consider symmetrical distributions with heavier tails. In this work, we studied the linear spatial model with normal/independent distribution, which is an alternative to the Gaussian spacial model. In this class, more robust distributions are included, such as t-Student and slash. We developed estimation methods by maximizing the likelihood function and EM algorithm. We also performed simulations under different scenarios to evaluate the performance of the algorithms. The construction of thematic maps was performed using the kriging technique adapted to the proposed model. Using the local influence method proposed by Cook (1986), we show a diagnostic study of the linear spacial model with normal/independent distribution. We conclude with an illustration of the proposed methodology with a data set of daily precipitation in Switzerland after the Chernobyl disaster occurred in 1986.