Construction of kriging prediction intervals for non-Gaussian spatial processes
(English)Manuscript (preprint) (Other academic)
In this article, we compare three methods to construct prediction intervals for the value of a stationary process, based on plug-in ordinary kriging predictors. Ordinary kriging is a widely used method for prediction that, given observations of a (spatial) process, forms the best linear unbiased predictor of the process at a new location. Construction of prediction intervals for the value of interest based on ordinary kriging predictors typically rely on Gaussian assumptions. Special attention is here given to non-Gaussian processes, where construction of such intervals is less straightforward. Methods based on asymptotic normality, Gaussian transformations and semiparametric bootstrap are compared on simulated and real data. The study suggests that the semiparametric method (that does not rely on distributional assumptions) is robust and is to be recommended for non-Gaussian processes. For practitioners the semiparametric method is an attractive alternative since the method can be used without spcifying a link function or making distributional assumptions.
Probability Theory and Statistics
Research subject Mathematical Statistics
IdentifiersURN: urn:nbn:se:umu:diva-53277OAI: oai:DiVA.org:umu-53277DiVA: diva2:511106